Loading...
1/*
2 * linux/net/sunrpc/svc.c
3 *
4 * High-level RPC service routines
5 *
6 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
7 *
8 * Multiple threads pools and NUMAisation
9 * Copyright (c) 2006 Silicon Graphics, Inc.
10 * by Greg Banks <gnb@melbourne.sgi.com>
11 */
12
13#include <linux/linkage.h>
14#include <linux/sched.h>
15#include <linux/errno.h>
16#include <linux/net.h>
17#include <linux/in.h>
18#include <linux/mm.h>
19#include <linux/interrupt.h>
20#include <linux/module.h>
21#include <linux/kthread.h>
22#include <linux/slab.h>
23
24#include <linux/sunrpc/types.h>
25#include <linux/sunrpc/xdr.h>
26#include <linux/sunrpc/stats.h>
27#include <linux/sunrpc/svcsock.h>
28#include <linux/sunrpc/clnt.h>
29#include <linux/sunrpc/bc_xprt.h>
30
31#include <trace/events/sunrpc.h>
32
33#define RPCDBG_FACILITY RPCDBG_SVCDSP
34
35static void svc_unregister(const struct svc_serv *serv, struct net *net);
36
37#define svc_serv_is_pooled(serv) ((serv)->sv_ops->svo_function)
38
39#define SVC_POOL_DEFAULT SVC_POOL_GLOBAL
40
41/*
42 * Structure for mapping cpus to pools and vice versa.
43 * Setup once during sunrpc initialisation.
44 */
45struct svc_pool_map svc_pool_map = {
46 .mode = SVC_POOL_DEFAULT
47};
48EXPORT_SYMBOL_GPL(svc_pool_map);
49
50static DEFINE_MUTEX(svc_pool_map_mutex);/* protects svc_pool_map.count only */
51
52static int
53param_set_pool_mode(const char *val, struct kernel_param *kp)
54{
55 int *ip = (int *)kp->arg;
56 struct svc_pool_map *m = &svc_pool_map;
57 int err;
58
59 mutex_lock(&svc_pool_map_mutex);
60
61 err = -EBUSY;
62 if (m->count)
63 goto out;
64
65 err = 0;
66 if (!strncmp(val, "auto", 4))
67 *ip = SVC_POOL_AUTO;
68 else if (!strncmp(val, "global", 6))
69 *ip = SVC_POOL_GLOBAL;
70 else if (!strncmp(val, "percpu", 6))
71 *ip = SVC_POOL_PERCPU;
72 else if (!strncmp(val, "pernode", 7))
73 *ip = SVC_POOL_PERNODE;
74 else
75 err = -EINVAL;
76
77out:
78 mutex_unlock(&svc_pool_map_mutex);
79 return err;
80}
81
82static int
83param_get_pool_mode(char *buf, struct kernel_param *kp)
84{
85 int *ip = (int *)kp->arg;
86
87 switch (*ip)
88 {
89 case SVC_POOL_AUTO:
90 return strlcpy(buf, "auto", 20);
91 case SVC_POOL_GLOBAL:
92 return strlcpy(buf, "global", 20);
93 case SVC_POOL_PERCPU:
94 return strlcpy(buf, "percpu", 20);
95 case SVC_POOL_PERNODE:
96 return strlcpy(buf, "pernode", 20);
97 default:
98 return sprintf(buf, "%d", *ip);
99 }
100}
101
102module_param_call(pool_mode, param_set_pool_mode, param_get_pool_mode,
103 &svc_pool_map.mode, 0644);
104
105/*
106 * Detect best pool mapping mode heuristically,
107 * according to the machine's topology.
108 */
109static int
110svc_pool_map_choose_mode(void)
111{
112 unsigned int node;
113
114 if (nr_online_nodes > 1) {
115 /*
116 * Actually have multiple NUMA nodes,
117 * so split pools on NUMA node boundaries
118 */
119 return SVC_POOL_PERNODE;
120 }
121
122 node = first_online_node;
123 if (nr_cpus_node(node) > 2) {
124 /*
125 * Non-trivial SMP, or CONFIG_NUMA on
126 * non-NUMA hardware, e.g. with a generic
127 * x86_64 kernel on Xeons. In this case we
128 * want to divide the pools on cpu boundaries.
129 */
130 return SVC_POOL_PERCPU;
131 }
132
133 /* default: one global pool */
134 return SVC_POOL_GLOBAL;
135}
136
137/*
138 * Allocate the to_pool[] and pool_to[] arrays.
139 * Returns 0 on success or an errno.
140 */
141static int
142svc_pool_map_alloc_arrays(struct svc_pool_map *m, unsigned int maxpools)
143{
144 m->to_pool = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
145 if (!m->to_pool)
146 goto fail;
147 m->pool_to = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
148 if (!m->pool_to)
149 goto fail_free;
150
151 return 0;
152
153fail_free:
154 kfree(m->to_pool);
155 m->to_pool = NULL;
156fail:
157 return -ENOMEM;
158}
159
160/*
161 * Initialise the pool map for SVC_POOL_PERCPU mode.
162 * Returns number of pools or <0 on error.
163 */
164static int
165svc_pool_map_init_percpu(struct svc_pool_map *m)
166{
167 unsigned int maxpools = nr_cpu_ids;
168 unsigned int pidx = 0;
169 unsigned int cpu;
170 int err;
171
172 err = svc_pool_map_alloc_arrays(m, maxpools);
173 if (err)
174 return err;
175
176 for_each_online_cpu(cpu) {
177 BUG_ON(pidx >= maxpools);
178 m->to_pool[cpu] = pidx;
179 m->pool_to[pidx] = cpu;
180 pidx++;
181 }
182 /* cpus brought online later all get mapped to pool0, sorry */
183
184 return pidx;
185};
186
187
188/*
189 * Initialise the pool map for SVC_POOL_PERNODE mode.
190 * Returns number of pools or <0 on error.
191 */
192static int
193svc_pool_map_init_pernode(struct svc_pool_map *m)
194{
195 unsigned int maxpools = nr_node_ids;
196 unsigned int pidx = 0;
197 unsigned int node;
198 int err;
199
200 err = svc_pool_map_alloc_arrays(m, maxpools);
201 if (err)
202 return err;
203
204 for_each_node_with_cpus(node) {
205 /* some architectures (e.g. SN2) have cpuless nodes */
206 BUG_ON(pidx > maxpools);
207 m->to_pool[node] = pidx;
208 m->pool_to[pidx] = node;
209 pidx++;
210 }
211 /* nodes brought online later all get mapped to pool0, sorry */
212
213 return pidx;
214}
215
216
217/*
218 * Add a reference to the global map of cpus to pools (and
219 * vice versa). Initialise the map if we're the first user.
220 * Returns the number of pools.
221 */
222unsigned int
223svc_pool_map_get(void)
224{
225 struct svc_pool_map *m = &svc_pool_map;
226 int npools = -1;
227
228 mutex_lock(&svc_pool_map_mutex);
229
230 if (m->count++) {
231 mutex_unlock(&svc_pool_map_mutex);
232 return m->npools;
233 }
234
235 if (m->mode == SVC_POOL_AUTO)
236 m->mode = svc_pool_map_choose_mode();
237
238 switch (m->mode) {
239 case SVC_POOL_PERCPU:
240 npools = svc_pool_map_init_percpu(m);
241 break;
242 case SVC_POOL_PERNODE:
243 npools = svc_pool_map_init_pernode(m);
244 break;
245 }
246
247 if (npools < 0) {
248 /* default, or memory allocation failure */
249 npools = 1;
250 m->mode = SVC_POOL_GLOBAL;
251 }
252 m->npools = npools;
253
254 mutex_unlock(&svc_pool_map_mutex);
255 return m->npools;
256}
257EXPORT_SYMBOL_GPL(svc_pool_map_get);
258
259/*
260 * Drop a reference to the global map of cpus to pools.
261 * When the last reference is dropped, the map data is
262 * freed; this allows the sysadmin to change the pool
263 * mode using the pool_mode module option without
264 * rebooting or re-loading sunrpc.ko.
265 */
266void
267svc_pool_map_put(void)
268{
269 struct svc_pool_map *m = &svc_pool_map;
270
271 mutex_lock(&svc_pool_map_mutex);
272
273 if (!--m->count) {
274 kfree(m->to_pool);
275 m->to_pool = NULL;
276 kfree(m->pool_to);
277 m->pool_to = NULL;
278 m->npools = 0;
279 }
280
281 mutex_unlock(&svc_pool_map_mutex);
282}
283EXPORT_SYMBOL_GPL(svc_pool_map_put);
284
285static int svc_pool_map_get_node(unsigned int pidx)
286{
287 const struct svc_pool_map *m = &svc_pool_map;
288
289 if (m->count) {
290 if (m->mode == SVC_POOL_PERCPU)
291 return cpu_to_node(m->pool_to[pidx]);
292 if (m->mode == SVC_POOL_PERNODE)
293 return m->pool_to[pidx];
294 }
295 return NUMA_NO_NODE;
296}
297/*
298 * Set the given thread's cpus_allowed mask so that it
299 * will only run on cpus in the given pool.
300 */
301static inline void
302svc_pool_map_set_cpumask(struct task_struct *task, unsigned int pidx)
303{
304 struct svc_pool_map *m = &svc_pool_map;
305 unsigned int node = m->pool_to[pidx];
306
307 /*
308 * The caller checks for sv_nrpools > 1, which
309 * implies that we've been initialized.
310 */
311 WARN_ON_ONCE(m->count == 0);
312 if (m->count == 0)
313 return;
314
315 switch (m->mode) {
316 case SVC_POOL_PERCPU:
317 {
318 set_cpus_allowed_ptr(task, cpumask_of(node));
319 break;
320 }
321 case SVC_POOL_PERNODE:
322 {
323 set_cpus_allowed_ptr(task, cpumask_of_node(node));
324 break;
325 }
326 }
327}
328
329/*
330 * Use the mapping mode to choose a pool for a given CPU.
331 * Used when enqueueing an incoming RPC. Always returns
332 * a non-NULL pool pointer.
333 */
334struct svc_pool *
335svc_pool_for_cpu(struct svc_serv *serv, int cpu)
336{
337 struct svc_pool_map *m = &svc_pool_map;
338 unsigned int pidx = 0;
339
340 /*
341 * An uninitialised map happens in a pure client when
342 * lockd is brought up, so silently treat it the
343 * same as SVC_POOL_GLOBAL.
344 */
345 if (svc_serv_is_pooled(serv)) {
346 switch (m->mode) {
347 case SVC_POOL_PERCPU:
348 pidx = m->to_pool[cpu];
349 break;
350 case SVC_POOL_PERNODE:
351 pidx = m->to_pool[cpu_to_node(cpu)];
352 break;
353 }
354 }
355 return &serv->sv_pools[pidx % serv->sv_nrpools];
356}
357
358int svc_rpcb_setup(struct svc_serv *serv, struct net *net)
359{
360 int err;
361
362 err = rpcb_create_local(net);
363 if (err)
364 return err;
365
366 /* Remove any stale portmap registrations */
367 svc_unregister(serv, net);
368 return 0;
369}
370EXPORT_SYMBOL_GPL(svc_rpcb_setup);
371
372void svc_rpcb_cleanup(struct svc_serv *serv, struct net *net)
373{
374 svc_unregister(serv, net);
375 rpcb_put_local(net);
376}
377EXPORT_SYMBOL_GPL(svc_rpcb_cleanup);
378
379static int svc_uses_rpcbind(struct svc_serv *serv)
380{
381 struct svc_program *progp;
382 unsigned int i;
383
384 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
385 for (i = 0; i < progp->pg_nvers; i++) {
386 if (progp->pg_vers[i] == NULL)
387 continue;
388 if (progp->pg_vers[i]->vs_hidden == 0)
389 return 1;
390 }
391 }
392
393 return 0;
394}
395
396int svc_bind(struct svc_serv *serv, struct net *net)
397{
398 if (!svc_uses_rpcbind(serv))
399 return 0;
400 return svc_rpcb_setup(serv, net);
401}
402EXPORT_SYMBOL_GPL(svc_bind);
403
404/*
405 * Create an RPC service
406 */
407static struct svc_serv *
408__svc_create(struct svc_program *prog, unsigned int bufsize, int npools,
409 struct svc_serv_ops *ops)
410{
411 struct svc_serv *serv;
412 unsigned int vers;
413 unsigned int xdrsize;
414 unsigned int i;
415
416 if (!(serv = kzalloc(sizeof(*serv), GFP_KERNEL)))
417 return NULL;
418 serv->sv_name = prog->pg_name;
419 serv->sv_program = prog;
420 serv->sv_nrthreads = 1;
421 serv->sv_stats = prog->pg_stats;
422 if (bufsize > RPCSVC_MAXPAYLOAD)
423 bufsize = RPCSVC_MAXPAYLOAD;
424 serv->sv_max_payload = bufsize? bufsize : 4096;
425 serv->sv_max_mesg = roundup(serv->sv_max_payload + PAGE_SIZE, PAGE_SIZE);
426 serv->sv_ops = ops;
427 xdrsize = 0;
428 while (prog) {
429 prog->pg_lovers = prog->pg_nvers-1;
430 for (vers=0; vers<prog->pg_nvers ; vers++)
431 if (prog->pg_vers[vers]) {
432 prog->pg_hivers = vers;
433 if (prog->pg_lovers > vers)
434 prog->pg_lovers = vers;
435 if (prog->pg_vers[vers]->vs_xdrsize > xdrsize)
436 xdrsize = prog->pg_vers[vers]->vs_xdrsize;
437 }
438 prog = prog->pg_next;
439 }
440 serv->sv_xdrsize = xdrsize;
441 INIT_LIST_HEAD(&serv->sv_tempsocks);
442 INIT_LIST_HEAD(&serv->sv_permsocks);
443 init_timer(&serv->sv_temptimer);
444 spin_lock_init(&serv->sv_lock);
445
446 serv->sv_nrpools = npools;
447 serv->sv_pools =
448 kcalloc(serv->sv_nrpools, sizeof(struct svc_pool),
449 GFP_KERNEL);
450 if (!serv->sv_pools) {
451 kfree(serv);
452 return NULL;
453 }
454
455 for (i = 0; i < serv->sv_nrpools; i++) {
456 struct svc_pool *pool = &serv->sv_pools[i];
457
458 dprintk("svc: initialising pool %u for %s\n",
459 i, serv->sv_name);
460
461 pool->sp_id = i;
462 INIT_LIST_HEAD(&pool->sp_sockets);
463 INIT_LIST_HEAD(&pool->sp_all_threads);
464 spin_lock_init(&pool->sp_lock);
465 }
466
467 return serv;
468}
469
470struct svc_serv *
471svc_create(struct svc_program *prog, unsigned int bufsize,
472 struct svc_serv_ops *ops)
473{
474 return __svc_create(prog, bufsize, /*npools*/1, ops);
475}
476EXPORT_SYMBOL_GPL(svc_create);
477
478struct svc_serv *
479svc_create_pooled(struct svc_program *prog, unsigned int bufsize,
480 struct svc_serv_ops *ops)
481{
482 struct svc_serv *serv;
483 unsigned int npools = svc_pool_map_get();
484
485 serv = __svc_create(prog, bufsize, npools, ops);
486 if (!serv)
487 goto out_err;
488 return serv;
489out_err:
490 svc_pool_map_put();
491 return NULL;
492}
493EXPORT_SYMBOL_GPL(svc_create_pooled);
494
495void svc_shutdown_net(struct svc_serv *serv, struct net *net)
496{
497 svc_close_net(serv, net);
498
499 if (serv->sv_ops->svo_shutdown)
500 serv->sv_ops->svo_shutdown(serv, net);
501}
502EXPORT_SYMBOL_GPL(svc_shutdown_net);
503
504/*
505 * Destroy an RPC service. Should be called with appropriate locking to
506 * protect the sv_nrthreads, sv_permsocks and sv_tempsocks.
507 */
508void
509svc_destroy(struct svc_serv *serv)
510{
511 dprintk("svc: svc_destroy(%s, %d)\n",
512 serv->sv_program->pg_name,
513 serv->sv_nrthreads);
514
515 if (serv->sv_nrthreads) {
516 if (--(serv->sv_nrthreads) != 0) {
517 svc_sock_update_bufs(serv);
518 return;
519 }
520 } else
521 printk("svc_destroy: no threads for serv=%p!\n", serv);
522
523 del_timer_sync(&serv->sv_temptimer);
524
525 /*
526 * The last user is gone and thus all sockets have to be destroyed to
527 * the point. Check this.
528 */
529 BUG_ON(!list_empty(&serv->sv_permsocks));
530 BUG_ON(!list_empty(&serv->sv_tempsocks));
531
532 cache_clean_deferred(serv);
533
534 if (svc_serv_is_pooled(serv))
535 svc_pool_map_put();
536
537 kfree(serv->sv_pools);
538 kfree(serv);
539}
540EXPORT_SYMBOL_GPL(svc_destroy);
541
542/*
543 * Allocate an RPC server's buffer space.
544 * We allocate pages and place them in rq_argpages.
545 */
546static int
547svc_init_buffer(struct svc_rqst *rqstp, unsigned int size, int node)
548{
549 unsigned int pages, arghi;
550
551 /* bc_xprt uses fore channel allocated buffers */
552 if (svc_is_backchannel(rqstp))
553 return 1;
554
555 pages = size / PAGE_SIZE + 1; /* extra page as we hold both request and reply.
556 * We assume one is at most one page
557 */
558 arghi = 0;
559 WARN_ON_ONCE(pages > RPCSVC_MAXPAGES);
560 if (pages > RPCSVC_MAXPAGES)
561 pages = RPCSVC_MAXPAGES;
562 while (pages) {
563 struct page *p = alloc_pages_node(node, GFP_KERNEL, 0);
564 if (!p)
565 break;
566 rqstp->rq_pages[arghi++] = p;
567 pages--;
568 }
569 return pages == 0;
570}
571
572/*
573 * Release an RPC server buffer
574 */
575static void
576svc_release_buffer(struct svc_rqst *rqstp)
577{
578 unsigned int i;
579
580 for (i = 0; i < ARRAY_SIZE(rqstp->rq_pages); i++)
581 if (rqstp->rq_pages[i])
582 put_page(rqstp->rq_pages[i]);
583}
584
585struct svc_rqst *
586svc_rqst_alloc(struct svc_serv *serv, struct svc_pool *pool, int node)
587{
588 struct svc_rqst *rqstp;
589
590 rqstp = kzalloc_node(sizeof(*rqstp), GFP_KERNEL, node);
591 if (!rqstp)
592 return rqstp;
593
594 __set_bit(RQ_BUSY, &rqstp->rq_flags);
595 spin_lock_init(&rqstp->rq_lock);
596 rqstp->rq_server = serv;
597 rqstp->rq_pool = pool;
598
599 rqstp->rq_argp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
600 if (!rqstp->rq_argp)
601 goto out_enomem;
602
603 rqstp->rq_resp = kmalloc_node(serv->sv_xdrsize, GFP_KERNEL, node);
604 if (!rqstp->rq_resp)
605 goto out_enomem;
606
607 if (!svc_init_buffer(rqstp, serv->sv_max_mesg, node))
608 goto out_enomem;
609
610 return rqstp;
611out_enomem:
612 svc_rqst_free(rqstp);
613 return NULL;
614}
615EXPORT_SYMBOL_GPL(svc_rqst_alloc);
616
617struct svc_rqst *
618svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool, int node)
619{
620 struct svc_rqst *rqstp;
621
622 rqstp = svc_rqst_alloc(serv, pool, node);
623 if (!rqstp)
624 return ERR_PTR(-ENOMEM);
625
626 serv->sv_nrthreads++;
627 spin_lock_bh(&pool->sp_lock);
628 pool->sp_nrthreads++;
629 list_add_rcu(&rqstp->rq_all, &pool->sp_all_threads);
630 spin_unlock_bh(&pool->sp_lock);
631 return rqstp;
632}
633EXPORT_SYMBOL_GPL(svc_prepare_thread);
634
635/*
636 * Choose a pool in which to create a new thread, for svc_set_num_threads
637 */
638static inline struct svc_pool *
639choose_pool(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
640{
641 if (pool != NULL)
642 return pool;
643
644 return &serv->sv_pools[(*state)++ % serv->sv_nrpools];
645}
646
647/*
648 * Choose a thread to kill, for svc_set_num_threads
649 */
650static inline struct task_struct *
651choose_victim(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
652{
653 unsigned int i;
654 struct task_struct *task = NULL;
655
656 if (pool != NULL) {
657 spin_lock_bh(&pool->sp_lock);
658 } else {
659 /* choose a pool in round-robin fashion */
660 for (i = 0; i < serv->sv_nrpools; i++) {
661 pool = &serv->sv_pools[--(*state) % serv->sv_nrpools];
662 spin_lock_bh(&pool->sp_lock);
663 if (!list_empty(&pool->sp_all_threads))
664 goto found_pool;
665 spin_unlock_bh(&pool->sp_lock);
666 }
667 return NULL;
668 }
669
670found_pool:
671 if (!list_empty(&pool->sp_all_threads)) {
672 struct svc_rqst *rqstp;
673
674 /*
675 * Remove from the pool->sp_all_threads list
676 * so we don't try to kill it again.
677 */
678 rqstp = list_entry(pool->sp_all_threads.next, struct svc_rqst, rq_all);
679 set_bit(RQ_VICTIM, &rqstp->rq_flags);
680 list_del_rcu(&rqstp->rq_all);
681 task = rqstp->rq_task;
682 }
683 spin_unlock_bh(&pool->sp_lock);
684
685 return task;
686}
687
688/*
689 * Create or destroy enough new threads to make the number
690 * of threads the given number. If `pool' is non-NULL, applies
691 * only to threads in that pool, otherwise round-robins between
692 * all pools. Caller must ensure that mutual exclusion between this and
693 * server startup or shutdown.
694 *
695 * Destroying threads relies on the service threads filling in
696 * rqstp->rq_task, which only the nfs ones do. Assumes the serv
697 * has been created using svc_create_pooled().
698 *
699 * Based on code that used to be in nfsd_svc() but tweaked
700 * to be pool-aware.
701 */
702int
703svc_set_num_threads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
704{
705 struct svc_rqst *rqstp;
706 struct task_struct *task;
707 struct svc_pool *chosen_pool;
708 int error = 0;
709 unsigned int state = serv->sv_nrthreads-1;
710 int node;
711
712 if (pool == NULL) {
713 /* The -1 assumes caller has done a svc_get() */
714 nrservs -= (serv->sv_nrthreads-1);
715 } else {
716 spin_lock_bh(&pool->sp_lock);
717 nrservs -= pool->sp_nrthreads;
718 spin_unlock_bh(&pool->sp_lock);
719 }
720
721 /* create new threads */
722 while (nrservs > 0) {
723 nrservs--;
724 chosen_pool = choose_pool(serv, pool, &state);
725
726 node = svc_pool_map_get_node(chosen_pool->sp_id);
727 rqstp = svc_prepare_thread(serv, chosen_pool, node);
728 if (IS_ERR(rqstp)) {
729 error = PTR_ERR(rqstp);
730 break;
731 }
732
733 __module_get(serv->sv_ops->svo_module);
734 task = kthread_create_on_node(serv->sv_ops->svo_function, rqstp,
735 node, "%s", serv->sv_name);
736 if (IS_ERR(task)) {
737 error = PTR_ERR(task);
738 module_put(serv->sv_ops->svo_module);
739 svc_exit_thread(rqstp);
740 break;
741 }
742
743 rqstp->rq_task = task;
744 if (serv->sv_nrpools > 1)
745 svc_pool_map_set_cpumask(task, chosen_pool->sp_id);
746
747 svc_sock_update_bufs(serv);
748 wake_up_process(task);
749 }
750 /* destroy old threads */
751 while (nrservs < 0 &&
752 (task = choose_victim(serv, pool, &state)) != NULL) {
753 send_sig(SIGINT, task, 1);
754 nrservs++;
755 }
756
757 return error;
758}
759EXPORT_SYMBOL_GPL(svc_set_num_threads);
760
761/*
762 * Called from a server thread as it's exiting. Caller must hold the "service
763 * mutex" for the service.
764 */
765void
766svc_rqst_free(struct svc_rqst *rqstp)
767{
768 svc_release_buffer(rqstp);
769 kfree(rqstp->rq_resp);
770 kfree(rqstp->rq_argp);
771 kfree(rqstp->rq_auth_data);
772 kfree_rcu(rqstp, rq_rcu_head);
773}
774EXPORT_SYMBOL_GPL(svc_rqst_free);
775
776void
777svc_exit_thread(struct svc_rqst *rqstp)
778{
779 struct svc_serv *serv = rqstp->rq_server;
780 struct svc_pool *pool = rqstp->rq_pool;
781
782 spin_lock_bh(&pool->sp_lock);
783 pool->sp_nrthreads--;
784 if (!test_and_set_bit(RQ_VICTIM, &rqstp->rq_flags))
785 list_del_rcu(&rqstp->rq_all);
786 spin_unlock_bh(&pool->sp_lock);
787
788 svc_rqst_free(rqstp);
789
790 /* Release the server */
791 if (serv)
792 svc_destroy(serv);
793}
794EXPORT_SYMBOL_GPL(svc_exit_thread);
795
796/*
797 * Register an "inet" protocol family netid with the local
798 * rpcbind daemon via an rpcbind v4 SET request.
799 *
800 * No netconfig infrastructure is available in the kernel, so
801 * we map IP_ protocol numbers to netids by hand.
802 *
803 * Returns zero on success; a negative errno value is returned
804 * if any error occurs.
805 */
806static int __svc_rpcb_register4(struct net *net, const u32 program,
807 const u32 version,
808 const unsigned short protocol,
809 const unsigned short port)
810{
811 const struct sockaddr_in sin = {
812 .sin_family = AF_INET,
813 .sin_addr.s_addr = htonl(INADDR_ANY),
814 .sin_port = htons(port),
815 };
816 const char *netid;
817 int error;
818
819 switch (protocol) {
820 case IPPROTO_UDP:
821 netid = RPCBIND_NETID_UDP;
822 break;
823 case IPPROTO_TCP:
824 netid = RPCBIND_NETID_TCP;
825 break;
826 default:
827 return -ENOPROTOOPT;
828 }
829
830 error = rpcb_v4_register(net, program, version,
831 (const struct sockaddr *)&sin, netid);
832
833 /*
834 * User space didn't support rpcbind v4, so retry this
835 * registration request with the legacy rpcbind v2 protocol.
836 */
837 if (error == -EPROTONOSUPPORT)
838 error = rpcb_register(net, program, version, protocol, port);
839
840 return error;
841}
842
843#if IS_ENABLED(CONFIG_IPV6)
844/*
845 * Register an "inet6" protocol family netid with the local
846 * rpcbind daemon via an rpcbind v4 SET request.
847 *
848 * No netconfig infrastructure is available in the kernel, so
849 * we map IP_ protocol numbers to netids by hand.
850 *
851 * Returns zero on success; a negative errno value is returned
852 * if any error occurs.
853 */
854static int __svc_rpcb_register6(struct net *net, const u32 program,
855 const u32 version,
856 const unsigned short protocol,
857 const unsigned short port)
858{
859 const struct sockaddr_in6 sin6 = {
860 .sin6_family = AF_INET6,
861 .sin6_addr = IN6ADDR_ANY_INIT,
862 .sin6_port = htons(port),
863 };
864 const char *netid;
865 int error;
866
867 switch (protocol) {
868 case IPPROTO_UDP:
869 netid = RPCBIND_NETID_UDP6;
870 break;
871 case IPPROTO_TCP:
872 netid = RPCBIND_NETID_TCP6;
873 break;
874 default:
875 return -ENOPROTOOPT;
876 }
877
878 error = rpcb_v4_register(net, program, version,
879 (const struct sockaddr *)&sin6, netid);
880
881 /*
882 * User space didn't support rpcbind version 4, so we won't
883 * use a PF_INET6 listener.
884 */
885 if (error == -EPROTONOSUPPORT)
886 error = -EAFNOSUPPORT;
887
888 return error;
889}
890#endif /* IS_ENABLED(CONFIG_IPV6) */
891
892/*
893 * Register a kernel RPC service via rpcbind version 4.
894 *
895 * Returns zero on success; a negative errno value is returned
896 * if any error occurs.
897 */
898static int __svc_register(struct net *net, const char *progname,
899 const u32 program, const u32 version,
900 const int family,
901 const unsigned short protocol,
902 const unsigned short port)
903{
904 int error = -EAFNOSUPPORT;
905
906 switch (family) {
907 case PF_INET:
908 error = __svc_rpcb_register4(net, program, version,
909 protocol, port);
910 break;
911#if IS_ENABLED(CONFIG_IPV6)
912 case PF_INET6:
913 error = __svc_rpcb_register6(net, program, version,
914 protocol, port);
915#endif
916 }
917
918 return error;
919}
920
921/**
922 * svc_register - register an RPC service with the local portmapper
923 * @serv: svc_serv struct for the service to register
924 * @net: net namespace for the service to register
925 * @family: protocol family of service's listener socket
926 * @proto: transport protocol number to advertise
927 * @port: port to advertise
928 *
929 * Service is registered for any address in the passed-in protocol family
930 */
931int svc_register(const struct svc_serv *serv, struct net *net,
932 const int family, const unsigned short proto,
933 const unsigned short port)
934{
935 struct svc_program *progp;
936 struct svc_version *vers;
937 unsigned int i;
938 int error = 0;
939
940 WARN_ON_ONCE(proto == 0 && port == 0);
941 if (proto == 0 && port == 0)
942 return -EINVAL;
943
944 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
945 for (i = 0; i < progp->pg_nvers; i++) {
946 vers = progp->pg_vers[i];
947 if (vers == NULL)
948 continue;
949
950 dprintk("svc: svc_register(%sv%d, %s, %u, %u)%s\n",
951 progp->pg_name,
952 i,
953 proto == IPPROTO_UDP? "udp" : "tcp",
954 port,
955 family,
956 vers->vs_hidden ?
957 " (but not telling portmap)" : "");
958
959 if (vers->vs_hidden)
960 continue;
961
962 error = __svc_register(net, progp->pg_name, progp->pg_prog,
963 i, family, proto, port);
964
965 if (vers->vs_rpcb_optnl) {
966 error = 0;
967 continue;
968 }
969
970 if (error < 0) {
971 printk(KERN_WARNING "svc: failed to register "
972 "%sv%u RPC service (errno %d).\n",
973 progp->pg_name, i, -error);
974 break;
975 }
976 }
977 }
978
979 return error;
980}
981
982/*
983 * If user space is running rpcbind, it should take the v4 UNSET
984 * and clear everything for this [program, version]. If user space
985 * is running portmap, it will reject the v4 UNSET, but won't have
986 * any "inet6" entries anyway. So a PMAP_UNSET should be sufficient
987 * in this case to clear all existing entries for [program, version].
988 */
989static void __svc_unregister(struct net *net, const u32 program, const u32 version,
990 const char *progname)
991{
992 int error;
993
994 error = rpcb_v4_register(net, program, version, NULL, "");
995
996 /*
997 * User space didn't support rpcbind v4, so retry this
998 * request with the legacy rpcbind v2 protocol.
999 */
1000 if (error == -EPROTONOSUPPORT)
1001 error = rpcb_register(net, program, version, 0, 0);
1002
1003 dprintk("svc: %s(%sv%u), error %d\n",
1004 __func__, progname, version, error);
1005}
1006
1007/*
1008 * All netids, bind addresses and ports registered for [program, version]
1009 * are removed from the local rpcbind database (if the service is not
1010 * hidden) to make way for a new instance of the service.
1011 *
1012 * The result of unregistration is reported via dprintk for those who want
1013 * verification of the result, but is otherwise not important.
1014 */
1015static void svc_unregister(const struct svc_serv *serv, struct net *net)
1016{
1017 struct svc_program *progp;
1018 unsigned long flags;
1019 unsigned int i;
1020
1021 clear_thread_flag(TIF_SIGPENDING);
1022
1023 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
1024 for (i = 0; i < progp->pg_nvers; i++) {
1025 if (progp->pg_vers[i] == NULL)
1026 continue;
1027 if (progp->pg_vers[i]->vs_hidden)
1028 continue;
1029
1030 dprintk("svc: attempting to unregister %sv%u\n",
1031 progp->pg_name, i);
1032 __svc_unregister(net, progp->pg_prog, i, progp->pg_name);
1033 }
1034 }
1035
1036 spin_lock_irqsave(¤t->sighand->siglock, flags);
1037 recalc_sigpending();
1038 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
1039}
1040
1041/*
1042 * dprintk the given error with the address of the client that caused it.
1043 */
1044#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
1045static __printf(2, 3)
1046void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...)
1047{
1048 struct va_format vaf;
1049 va_list args;
1050 char buf[RPC_MAX_ADDRBUFLEN];
1051
1052 va_start(args, fmt);
1053
1054 vaf.fmt = fmt;
1055 vaf.va = &args;
1056
1057 dprintk("svc: %s: %pV", svc_print_addr(rqstp, buf, sizeof(buf)), &vaf);
1058
1059 va_end(args);
1060}
1061#else
1062static __printf(2,3) void svc_printk(struct svc_rqst *rqstp, const char *fmt, ...) {}
1063#endif
1064
1065/*
1066 * Common routine for processing the RPC request.
1067 */
1068static int
1069svc_process_common(struct svc_rqst *rqstp, struct kvec *argv, struct kvec *resv)
1070{
1071 struct svc_program *progp;
1072 struct svc_version *versp = NULL; /* compiler food */
1073 struct svc_procedure *procp = NULL;
1074 struct svc_serv *serv = rqstp->rq_server;
1075 kxdrproc_t xdr;
1076 __be32 *statp;
1077 u32 prog, vers, proc;
1078 __be32 auth_stat, rpc_stat;
1079 int auth_res;
1080 __be32 *reply_statp;
1081
1082 rpc_stat = rpc_success;
1083
1084 if (argv->iov_len < 6*4)
1085 goto err_short_len;
1086
1087 /* Will be turned off only in gss privacy case: */
1088 set_bit(RQ_SPLICE_OK, &rqstp->rq_flags);
1089 /* Will be turned off only when NFSv4 Sessions are used */
1090 set_bit(RQ_USEDEFERRAL, &rqstp->rq_flags);
1091 clear_bit(RQ_DROPME, &rqstp->rq_flags);
1092
1093 /* Setup reply header */
1094 rqstp->rq_xprt->xpt_ops->xpo_prep_reply_hdr(rqstp);
1095
1096 svc_putu32(resv, rqstp->rq_xid);
1097
1098 vers = svc_getnl(argv);
1099
1100 /* First words of reply: */
1101 svc_putnl(resv, 1); /* REPLY */
1102
1103 if (vers != 2) /* RPC version number */
1104 goto err_bad_rpc;
1105
1106 /* Save position in case we later decide to reject: */
1107 reply_statp = resv->iov_base + resv->iov_len;
1108
1109 svc_putnl(resv, 0); /* ACCEPT */
1110
1111 rqstp->rq_prog = prog = svc_getnl(argv); /* program number */
1112 rqstp->rq_vers = vers = svc_getnl(argv); /* version number */
1113 rqstp->rq_proc = proc = svc_getnl(argv); /* procedure number */
1114
1115 for (progp = serv->sv_program; progp; progp = progp->pg_next)
1116 if (prog == progp->pg_prog)
1117 break;
1118
1119 /*
1120 * Decode auth data, and add verifier to reply buffer.
1121 * We do this before anything else in order to get a decent
1122 * auth verifier.
1123 */
1124 auth_res = svc_authenticate(rqstp, &auth_stat);
1125 /* Also give the program a chance to reject this call: */
1126 if (auth_res == SVC_OK && progp) {
1127 auth_stat = rpc_autherr_badcred;
1128 auth_res = progp->pg_authenticate(rqstp);
1129 }
1130 switch (auth_res) {
1131 case SVC_OK:
1132 break;
1133 case SVC_GARBAGE:
1134 goto err_garbage;
1135 case SVC_SYSERR:
1136 rpc_stat = rpc_system_err;
1137 goto err_bad;
1138 case SVC_DENIED:
1139 goto err_bad_auth;
1140 case SVC_CLOSE:
1141 if (test_bit(XPT_TEMP, &rqstp->rq_xprt->xpt_flags))
1142 svc_close_xprt(rqstp->rq_xprt);
1143 case SVC_DROP:
1144 goto dropit;
1145 case SVC_COMPLETE:
1146 goto sendit;
1147 }
1148
1149 if (progp == NULL)
1150 goto err_bad_prog;
1151
1152 if (vers >= progp->pg_nvers ||
1153 !(versp = progp->pg_vers[vers]))
1154 goto err_bad_vers;
1155
1156 procp = versp->vs_proc + proc;
1157 if (proc >= versp->vs_nproc || !procp->pc_func)
1158 goto err_bad_proc;
1159 rqstp->rq_procinfo = procp;
1160
1161 /* Syntactic check complete */
1162 serv->sv_stats->rpccnt++;
1163
1164 /* Build the reply header. */
1165 statp = resv->iov_base +resv->iov_len;
1166 svc_putnl(resv, RPC_SUCCESS);
1167
1168 /* Bump per-procedure stats counter */
1169 procp->pc_count++;
1170
1171 /* Initialize storage for argp and resp */
1172 memset(rqstp->rq_argp, 0, procp->pc_argsize);
1173 memset(rqstp->rq_resp, 0, procp->pc_ressize);
1174
1175 /* un-reserve some of the out-queue now that we have a
1176 * better idea of reply size
1177 */
1178 if (procp->pc_xdrressize)
1179 svc_reserve_auth(rqstp, procp->pc_xdrressize<<2);
1180
1181 /* Call the function that processes the request. */
1182 if (!versp->vs_dispatch) {
1183 /* Decode arguments */
1184 xdr = procp->pc_decode;
1185 if (xdr && !xdr(rqstp, argv->iov_base, rqstp->rq_argp))
1186 goto err_garbage;
1187
1188 *statp = procp->pc_func(rqstp, rqstp->rq_argp, rqstp->rq_resp);
1189
1190 /* Encode reply */
1191 if (test_bit(RQ_DROPME, &rqstp->rq_flags)) {
1192 if (procp->pc_release)
1193 procp->pc_release(rqstp, NULL, rqstp->rq_resp);
1194 goto dropit;
1195 }
1196 if (*statp == rpc_success &&
1197 (xdr = procp->pc_encode) &&
1198 !xdr(rqstp, resv->iov_base+resv->iov_len, rqstp->rq_resp)) {
1199 dprintk("svc: failed to encode reply\n");
1200 /* serv->sv_stats->rpcsystemerr++; */
1201 *statp = rpc_system_err;
1202 }
1203 } else {
1204 dprintk("svc: calling dispatcher\n");
1205 if (!versp->vs_dispatch(rqstp, statp)) {
1206 /* Release reply info */
1207 if (procp->pc_release)
1208 procp->pc_release(rqstp, NULL, rqstp->rq_resp);
1209 goto dropit;
1210 }
1211 }
1212
1213 /* Check RPC status result */
1214 if (*statp != rpc_success)
1215 resv->iov_len = ((void*)statp) - resv->iov_base + 4;
1216
1217 /* Release reply info */
1218 if (procp->pc_release)
1219 procp->pc_release(rqstp, NULL, rqstp->rq_resp);
1220
1221 if (procp->pc_encode == NULL)
1222 goto dropit;
1223
1224 sendit:
1225 if (svc_authorise(rqstp))
1226 goto dropit;
1227 return 1; /* Caller can now send it */
1228
1229 dropit:
1230 svc_authorise(rqstp); /* doesn't hurt to call this twice */
1231 dprintk("svc: svc_process dropit\n");
1232 return 0;
1233
1234err_short_len:
1235 svc_printk(rqstp, "short len %Zd, dropping request\n",
1236 argv->iov_len);
1237
1238 goto dropit; /* drop request */
1239
1240err_bad_rpc:
1241 serv->sv_stats->rpcbadfmt++;
1242 svc_putnl(resv, 1); /* REJECT */
1243 svc_putnl(resv, 0); /* RPC_MISMATCH */
1244 svc_putnl(resv, 2); /* Only RPCv2 supported */
1245 svc_putnl(resv, 2);
1246 goto sendit;
1247
1248err_bad_auth:
1249 dprintk("svc: authentication failed (%d)\n", ntohl(auth_stat));
1250 serv->sv_stats->rpcbadauth++;
1251 /* Restore write pointer to location of accept status: */
1252 xdr_ressize_check(rqstp, reply_statp);
1253 svc_putnl(resv, 1); /* REJECT */
1254 svc_putnl(resv, 1); /* AUTH_ERROR */
1255 svc_putnl(resv, ntohl(auth_stat)); /* status */
1256 goto sendit;
1257
1258err_bad_prog:
1259 dprintk("svc: unknown program %d\n", prog);
1260 serv->sv_stats->rpcbadfmt++;
1261 svc_putnl(resv, RPC_PROG_UNAVAIL);
1262 goto sendit;
1263
1264err_bad_vers:
1265 svc_printk(rqstp, "unknown version (%d for prog %d, %s)\n",
1266 vers, prog, progp->pg_name);
1267
1268 serv->sv_stats->rpcbadfmt++;
1269 svc_putnl(resv, RPC_PROG_MISMATCH);
1270 svc_putnl(resv, progp->pg_lovers);
1271 svc_putnl(resv, progp->pg_hivers);
1272 goto sendit;
1273
1274err_bad_proc:
1275 svc_printk(rqstp, "unknown procedure (%d)\n", proc);
1276
1277 serv->sv_stats->rpcbadfmt++;
1278 svc_putnl(resv, RPC_PROC_UNAVAIL);
1279 goto sendit;
1280
1281err_garbage:
1282 svc_printk(rqstp, "failed to decode args\n");
1283
1284 rpc_stat = rpc_garbage_args;
1285err_bad:
1286 serv->sv_stats->rpcbadfmt++;
1287 svc_putnl(resv, ntohl(rpc_stat));
1288 goto sendit;
1289}
1290
1291/*
1292 * Process the RPC request.
1293 */
1294int
1295svc_process(struct svc_rqst *rqstp)
1296{
1297 struct kvec *argv = &rqstp->rq_arg.head[0];
1298 struct kvec *resv = &rqstp->rq_res.head[0];
1299 struct svc_serv *serv = rqstp->rq_server;
1300 u32 dir;
1301
1302 /*
1303 * Setup response xdr_buf.
1304 * Initially it has just one page
1305 */
1306 rqstp->rq_next_page = &rqstp->rq_respages[1];
1307 resv->iov_base = page_address(rqstp->rq_respages[0]);
1308 resv->iov_len = 0;
1309 rqstp->rq_res.pages = rqstp->rq_respages + 1;
1310 rqstp->rq_res.len = 0;
1311 rqstp->rq_res.page_base = 0;
1312 rqstp->rq_res.page_len = 0;
1313 rqstp->rq_res.buflen = PAGE_SIZE;
1314 rqstp->rq_res.tail[0].iov_base = NULL;
1315 rqstp->rq_res.tail[0].iov_len = 0;
1316
1317 dir = svc_getnl(argv);
1318 if (dir != 0) {
1319 /* direction != CALL */
1320 svc_printk(rqstp, "bad direction %d, dropping request\n", dir);
1321 serv->sv_stats->rpcbadfmt++;
1322 goto out_drop;
1323 }
1324
1325 /* Returns 1 for send, 0 for drop */
1326 if (likely(svc_process_common(rqstp, argv, resv))) {
1327 int ret = svc_send(rqstp);
1328
1329 trace_svc_process(rqstp, ret);
1330 return ret;
1331 }
1332out_drop:
1333 trace_svc_process(rqstp, 0);
1334 svc_drop(rqstp);
1335 return 0;
1336}
1337EXPORT_SYMBOL_GPL(svc_process);
1338
1339#if defined(CONFIG_SUNRPC_BACKCHANNEL)
1340/*
1341 * Process a backchannel RPC request that arrived over an existing
1342 * outbound connection
1343 */
1344int
1345bc_svc_process(struct svc_serv *serv, struct rpc_rqst *req,
1346 struct svc_rqst *rqstp)
1347{
1348 struct kvec *argv = &rqstp->rq_arg.head[0];
1349 struct kvec *resv = &rqstp->rq_res.head[0];
1350 struct rpc_task *task;
1351 int proc_error;
1352 int error;
1353
1354 dprintk("svc: %s(%p)\n", __func__, req);
1355
1356 /* Build the svc_rqst used by the common processing routine */
1357 rqstp->rq_xprt = serv->sv_bc_xprt;
1358 rqstp->rq_xid = req->rq_xid;
1359 rqstp->rq_prot = req->rq_xprt->prot;
1360 rqstp->rq_server = serv;
1361
1362 rqstp->rq_addrlen = sizeof(req->rq_xprt->addr);
1363 memcpy(&rqstp->rq_addr, &req->rq_xprt->addr, rqstp->rq_addrlen);
1364 memcpy(&rqstp->rq_arg, &req->rq_rcv_buf, sizeof(rqstp->rq_arg));
1365 memcpy(&rqstp->rq_res, &req->rq_snd_buf, sizeof(rqstp->rq_res));
1366
1367 /* Adjust the argument buffer length */
1368 rqstp->rq_arg.len = req->rq_private_buf.len;
1369 if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len) {
1370 rqstp->rq_arg.head[0].iov_len = rqstp->rq_arg.len;
1371 rqstp->rq_arg.page_len = 0;
1372 } else if (rqstp->rq_arg.len <= rqstp->rq_arg.head[0].iov_len +
1373 rqstp->rq_arg.page_len)
1374 rqstp->rq_arg.page_len = rqstp->rq_arg.len -
1375 rqstp->rq_arg.head[0].iov_len;
1376 else
1377 rqstp->rq_arg.len = rqstp->rq_arg.head[0].iov_len +
1378 rqstp->rq_arg.page_len;
1379
1380 /* reset result send buffer "put" position */
1381 resv->iov_len = 0;
1382
1383 /*
1384 * Skip the next two words because they've already been
1385 * processed in the transport
1386 */
1387 svc_getu32(argv); /* XID */
1388 svc_getnl(argv); /* CALLDIR */
1389
1390 /* Parse and execute the bc call */
1391 proc_error = svc_process_common(rqstp, argv, resv);
1392
1393 atomic_inc(&req->rq_xprt->bc_free_slots);
1394 if (!proc_error) {
1395 /* Processing error: drop the request */
1396 xprt_free_bc_request(req);
1397 return 0;
1398 }
1399
1400 /* Finally, send the reply synchronously */
1401 memcpy(&req->rq_snd_buf, &rqstp->rq_res, sizeof(req->rq_snd_buf));
1402 task = rpc_run_bc_task(req);
1403 if (IS_ERR(task)) {
1404 error = PTR_ERR(task);
1405 goto out;
1406 }
1407
1408 WARN_ON_ONCE(atomic_read(&task->tk_count) != 1);
1409 error = task->tk_status;
1410 rpc_put_task(task);
1411
1412out:
1413 dprintk("svc: %s(), error=%d\n", __func__, error);
1414 return error;
1415}
1416EXPORT_SYMBOL_GPL(bc_svc_process);
1417#endif /* CONFIG_SUNRPC_BACKCHANNEL */
1418
1419/*
1420 * Return (transport-specific) limit on the rpc payload.
1421 */
1422u32 svc_max_payload(const struct svc_rqst *rqstp)
1423{
1424 u32 max = rqstp->rq_xprt->xpt_class->xcl_max_payload;
1425
1426 if (rqstp->rq_server->sv_max_payload < max)
1427 max = rqstp->rq_server->sv_max_payload;
1428 return max;
1429}
1430EXPORT_SYMBOL_GPL(svc_max_payload);
1/*
2 * linux/net/sunrpc/svc.c
3 *
4 * High-level RPC service routines
5 *
6 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
7 *
8 * Multiple threads pools and NUMAisation
9 * Copyright (c) 2006 Silicon Graphics, Inc.
10 * by Greg Banks <gnb@melbourne.sgi.com>
11 */
12
13#include <linux/linkage.h>
14#include <linux/sched.h>
15#include <linux/errno.h>
16#include <linux/net.h>
17#include <linux/in.h>
18#include <linux/mm.h>
19#include <linux/interrupt.h>
20#include <linux/module.h>
21#include <linux/kthread.h>
22#include <linux/slab.h>
23
24#include <linux/sunrpc/types.h>
25#include <linux/sunrpc/xdr.h>
26#include <linux/sunrpc/stats.h>
27#include <linux/sunrpc/svcsock.h>
28#include <linux/sunrpc/clnt.h>
29#include <linux/sunrpc/bc_xprt.h>
30
31#define RPCDBG_FACILITY RPCDBG_SVCDSP
32
33static void svc_unregister(const struct svc_serv *serv);
34
35#define svc_serv_is_pooled(serv) ((serv)->sv_function)
36
37/*
38 * Mode for mapping cpus to pools.
39 */
40enum {
41 SVC_POOL_AUTO = -1, /* choose one of the others */
42 SVC_POOL_GLOBAL, /* no mapping, just a single global pool
43 * (legacy & UP mode) */
44 SVC_POOL_PERCPU, /* one pool per cpu */
45 SVC_POOL_PERNODE /* one pool per numa node */
46};
47#define SVC_POOL_DEFAULT SVC_POOL_GLOBAL
48
49/*
50 * Structure for mapping cpus to pools and vice versa.
51 * Setup once during sunrpc initialisation.
52 */
53static struct svc_pool_map {
54 int count; /* How many svc_servs use us */
55 int mode; /* Note: int not enum to avoid
56 * warnings about "enumeration value
57 * not handled in switch" */
58 unsigned int npools;
59 unsigned int *pool_to; /* maps pool id to cpu or node */
60 unsigned int *to_pool; /* maps cpu or node to pool id */
61} svc_pool_map = {
62 .count = 0,
63 .mode = SVC_POOL_DEFAULT
64};
65static DEFINE_MUTEX(svc_pool_map_mutex);/* protects svc_pool_map.count only */
66
67static int
68param_set_pool_mode(const char *val, struct kernel_param *kp)
69{
70 int *ip = (int *)kp->arg;
71 struct svc_pool_map *m = &svc_pool_map;
72 int err;
73
74 mutex_lock(&svc_pool_map_mutex);
75
76 err = -EBUSY;
77 if (m->count)
78 goto out;
79
80 err = 0;
81 if (!strncmp(val, "auto", 4))
82 *ip = SVC_POOL_AUTO;
83 else if (!strncmp(val, "global", 6))
84 *ip = SVC_POOL_GLOBAL;
85 else if (!strncmp(val, "percpu", 6))
86 *ip = SVC_POOL_PERCPU;
87 else if (!strncmp(val, "pernode", 7))
88 *ip = SVC_POOL_PERNODE;
89 else
90 err = -EINVAL;
91
92out:
93 mutex_unlock(&svc_pool_map_mutex);
94 return err;
95}
96
97static int
98param_get_pool_mode(char *buf, struct kernel_param *kp)
99{
100 int *ip = (int *)kp->arg;
101
102 switch (*ip)
103 {
104 case SVC_POOL_AUTO:
105 return strlcpy(buf, "auto", 20);
106 case SVC_POOL_GLOBAL:
107 return strlcpy(buf, "global", 20);
108 case SVC_POOL_PERCPU:
109 return strlcpy(buf, "percpu", 20);
110 case SVC_POOL_PERNODE:
111 return strlcpy(buf, "pernode", 20);
112 default:
113 return sprintf(buf, "%d", *ip);
114 }
115}
116
117module_param_call(pool_mode, param_set_pool_mode, param_get_pool_mode,
118 &svc_pool_map.mode, 0644);
119
120/*
121 * Detect best pool mapping mode heuristically,
122 * according to the machine's topology.
123 */
124static int
125svc_pool_map_choose_mode(void)
126{
127 unsigned int node;
128
129 if (nr_online_nodes > 1) {
130 /*
131 * Actually have multiple NUMA nodes,
132 * so split pools on NUMA node boundaries
133 */
134 return SVC_POOL_PERNODE;
135 }
136
137 node = first_online_node;
138 if (nr_cpus_node(node) > 2) {
139 /*
140 * Non-trivial SMP, or CONFIG_NUMA on
141 * non-NUMA hardware, e.g. with a generic
142 * x86_64 kernel on Xeons. In this case we
143 * want to divide the pools on cpu boundaries.
144 */
145 return SVC_POOL_PERCPU;
146 }
147
148 /* default: one global pool */
149 return SVC_POOL_GLOBAL;
150}
151
152/*
153 * Allocate the to_pool[] and pool_to[] arrays.
154 * Returns 0 on success or an errno.
155 */
156static int
157svc_pool_map_alloc_arrays(struct svc_pool_map *m, unsigned int maxpools)
158{
159 m->to_pool = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
160 if (!m->to_pool)
161 goto fail;
162 m->pool_to = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
163 if (!m->pool_to)
164 goto fail_free;
165
166 return 0;
167
168fail_free:
169 kfree(m->to_pool);
170fail:
171 return -ENOMEM;
172}
173
174/*
175 * Initialise the pool map for SVC_POOL_PERCPU mode.
176 * Returns number of pools or <0 on error.
177 */
178static int
179svc_pool_map_init_percpu(struct svc_pool_map *m)
180{
181 unsigned int maxpools = nr_cpu_ids;
182 unsigned int pidx = 0;
183 unsigned int cpu;
184 int err;
185
186 err = svc_pool_map_alloc_arrays(m, maxpools);
187 if (err)
188 return err;
189
190 for_each_online_cpu(cpu) {
191 BUG_ON(pidx > maxpools);
192 m->to_pool[cpu] = pidx;
193 m->pool_to[pidx] = cpu;
194 pidx++;
195 }
196 /* cpus brought online later all get mapped to pool0, sorry */
197
198 return pidx;
199};
200
201
202/*
203 * Initialise the pool map for SVC_POOL_PERNODE mode.
204 * Returns number of pools or <0 on error.
205 */
206static int
207svc_pool_map_init_pernode(struct svc_pool_map *m)
208{
209 unsigned int maxpools = nr_node_ids;
210 unsigned int pidx = 0;
211 unsigned int node;
212 int err;
213
214 err = svc_pool_map_alloc_arrays(m, maxpools);
215 if (err)
216 return err;
217
218 for_each_node_with_cpus(node) {
219 /* some architectures (e.g. SN2) have cpuless nodes */
220 BUG_ON(pidx > maxpools);
221 m->to_pool[node] = pidx;
222 m->pool_to[pidx] = node;
223 pidx++;
224 }
225 /* nodes brought online later all get mapped to pool0, sorry */
226
227 return pidx;
228}
229
230
231/*
232 * Add a reference to the global map of cpus to pools (and
233 * vice versa). Initialise the map if we're the first user.
234 * Returns the number of pools.
235 */
236static unsigned int
237svc_pool_map_get(void)
238{
239 struct svc_pool_map *m = &svc_pool_map;
240 int npools = -1;
241
242 mutex_lock(&svc_pool_map_mutex);
243
244 if (m->count++) {
245 mutex_unlock(&svc_pool_map_mutex);
246 return m->npools;
247 }
248
249 if (m->mode == SVC_POOL_AUTO)
250 m->mode = svc_pool_map_choose_mode();
251
252 switch (m->mode) {
253 case SVC_POOL_PERCPU:
254 npools = svc_pool_map_init_percpu(m);
255 break;
256 case SVC_POOL_PERNODE:
257 npools = svc_pool_map_init_pernode(m);
258 break;
259 }
260
261 if (npools < 0) {
262 /* default, or memory allocation failure */
263 npools = 1;
264 m->mode = SVC_POOL_GLOBAL;
265 }
266 m->npools = npools;
267
268 mutex_unlock(&svc_pool_map_mutex);
269 return m->npools;
270}
271
272
273/*
274 * Drop a reference to the global map of cpus to pools.
275 * When the last reference is dropped, the map data is
276 * freed; this allows the sysadmin to change the pool
277 * mode using the pool_mode module option without
278 * rebooting or re-loading sunrpc.ko.
279 */
280static void
281svc_pool_map_put(void)
282{
283 struct svc_pool_map *m = &svc_pool_map;
284
285 mutex_lock(&svc_pool_map_mutex);
286
287 if (!--m->count) {
288 m->mode = SVC_POOL_DEFAULT;
289 kfree(m->to_pool);
290 kfree(m->pool_to);
291 m->npools = 0;
292 }
293
294 mutex_unlock(&svc_pool_map_mutex);
295}
296
297
298/*
299 * Set the given thread's cpus_allowed mask so that it
300 * will only run on cpus in the given pool.
301 */
302static inline void
303svc_pool_map_set_cpumask(struct task_struct *task, unsigned int pidx)
304{
305 struct svc_pool_map *m = &svc_pool_map;
306 unsigned int node = m->pool_to[pidx];
307
308 /*
309 * The caller checks for sv_nrpools > 1, which
310 * implies that we've been initialized.
311 */
312 BUG_ON(m->count == 0);
313
314 switch (m->mode) {
315 case SVC_POOL_PERCPU:
316 {
317 set_cpus_allowed_ptr(task, cpumask_of(node));
318 break;
319 }
320 case SVC_POOL_PERNODE:
321 {
322 set_cpus_allowed_ptr(task, cpumask_of_node(node));
323 break;
324 }
325 }
326}
327
328/*
329 * Use the mapping mode to choose a pool for a given CPU.
330 * Used when enqueueing an incoming RPC. Always returns
331 * a non-NULL pool pointer.
332 */
333struct svc_pool *
334svc_pool_for_cpu(struct svc_serv *serv, int cpu)
335{
336 struct svc_pool_map *m = &svc_pool_map;
337 unsigned int pidx = 0;
338
339 /*
340 * An uninitialised map happens in a pure client when
341 * lockd is brought up, so silently treat it the
342 * same as SVC_POOL_GLOBAL.
343 */
344 if (svc_serv_is_pooled(serv)) {
345 switch (m->mode) {
346 case SVC_POOL_PERCPU:
347 pidx = m->to_pool[cpu];
348 break;
349 case SVC_POOL_PERNODE:
350 pidx = m->to_pool[cpu_to_node(cpu)];
351 break;
352 }
353 }
354 return &serv->sv_pools[pidx % serv->sv_nrpools];
355}
356
357
358/*
359 * Create an RPC service
360 */
361static struct svc_serv *
362__svc_create(struct svc_program *prog, unsigned int bufsize, int npools,
363 void (*shutdown)(struct svc_serv *serv))
364{
365 struct svc_serv *serv;
366 unsigned int vers;
367 unsigned int xdrsize;
368 unsigned int i;
369
370 if (!(serv = kzalloc(sizeof(*serv), GFP_KERNEL)))
371 return NULL;
372 serv->sv_name = prog->pg_name;
373 serv->sv_program = prog;
374 serv->sv_nrthreads = 1;
375 serv->sv_stats = prog->pg_stats;
376 if (bufsize > RPCSVC_MAXPAYLOAD)
377 bufsize = RPCSVC_MAXPAYLOAD;
378 serv->sv_max_payload = bufsize? bufsize : 4096;
379 serv->sv_max_mesg = roundup(serv->sv_max_payload + PAGE_SIZE, PAGE_SIZE);
380 serv->sv_shutdown = shutdown;
381 xdrsize = 0;
382 while (prog) {
383 prog->pg_lovers = prog->pg_nvers-1;
384 for (vers=0; vers<prog->pg_nvers ; vers++)
385 if (prog->pg_vers[vers]) {
386 prog->pg_hivers = vers;
387 if (prog->pg_lovers > vers)
388 prog->pg_lovers = vers;
389 if (prog->pg_vers[vers]->vs_xdrsize > xdrsize)
390 xdrsize = prog->pg_vers[vers]->vs_xdrsize;
391 }
392 prog = prog->pg_next;
393 }
394 serv->sv_xdrsize = xdrsize;
395 INIT_LIST_HEAD(&serv->sv_tempsocks);
396 INIT_LIST_HEAD(&serv->sv_permsocks);
397 init_timer(&serv->sv_temptimer);
398 spin_lock_init(&serv->sv_lock);
399
400 serv->sv_nrpools = npools;
401 serv->sv_pools =
402 kcalloc(serv->sv_nrpools, sizeof(struct svc_pool),
403 GFP_KERNEL);
404 if (!serv->sv_pools) {
405 kfree(serv);
406 return NULL;
407 }
408
409 for (i = 0; i < serv->sv_nrpools; i++) {
410 struct svc_pool *pool = &serv->sv_pools[i];
411
412 dprintk("svc: initialising pool %u for %s\n",
413 i, serv->sv_name);
414
415 pool->sp_id = i;
416 INIT_LIST_HEAD(&pool->sp_threads);
417 INIT_LIST_HEAD(&pool->sp_sockets);
418 INIT_LIST_HEAD(&pool->sp_all_threads);
419 spin_lock_init(&pool->sp_lock);
420 }
421
422 /* Remove any stale portmap registrations */
423 svc_unregister(serv);
424
425 return serv;
426}
427
428struct svc_serv *
429svc_create(struct svc_program *prog, unsigned int bufsize,
430 void (*shutdown)(struct svc_serv *serv))
431{
432 return __svc_create(prog, bufsize, /*npools*/1, shutdown);
433}
434EXPORT_SYMBOL_GPL(svc_create);
435
436struct svc_serv *
437svc_create_pooled(struct svc_program *prog, unsigned int bufsize,
438 void (*shutdown)(struct svc_serv *serv),
439 svc_thread_fn func, struct module *mod)
440{
441 struct svc_serv *serv;
442 unsigned int npools = svc_pool_map_get();
443
444 serv = __svc_create(prog, bufsize, npools, shutdown);
445
446 if (serv != NULL) {
447 serv->sv_function = func;
448 serv->sv_module = mod;
449 }
450
451 return serv;
452}
453EXPORT_SYMBOL_GPL(svc_create_pooled);
454
455/*
456 * Destroy an RPC service. Should be called with appropriate locking to
457 * protect the sv_nrthreads, sv_permsocks and sv_tempsocks.
458 */
459void
460svc_destroy(struct svc_serv *serv)
461{
462 dprintk("svc: svc_destroy(%s, %d)\n",
463 serv->sv_program->pg_name,
464 serv->sv_nrthreads);
465
466 if (serv->sv_nrthreads) {
467 if (--(serv->sv_nrthreads) != 0) {
468 svc_sock_update_bufs(serv);
469 return;
470 }
471 } else
472 printk("svc_destroy: no threads for serv=%p!\n", serv);
473
474 del_timer_sync(&serv->sv_temptimer);
475
476 svc_close_all(&serv->sv_tempsocks);
477
478 if (serv->sv_shutdown)
479 serv->sv_shutdown(serv);
480
481 svc_close_all(&serv->sv_permsocks);
482
483 BUG_ON(!list_empty(&serv->sv_permsocks));
484 BUG_ON(!list_empty(&serv->sv_tempsocks));
485
486 cache_clean_deferred(serv);
487
488 if (svc_serv_is_pooled(serv))
489 svc_pool_map_put();
490
491 svc_unregister(serv);
492 kfree(serv->sv_pools);
493 kfree(serv);
494}
495EXPORT_SYMBOL_GPL(svc_destroy);
496
497/*
498 * Allocate an RPC server's buffer space.
499 * We allocate pages and place them in rq_argpages.
500 */
501static int
502svc_init_buffer(struct svc_rqst *rqstp, unsigned int size)
503{
504 unsigned int pages, arghi;
505
506 /* bc_xprt uses fore channel allocated buffers */
507 if (svc_is_backchannel(rqstp))
508 return 1;
509
510 pages = size / PAGE_SIZE + 1; /* extra page as we hold both request and reply.
511 * We assume one is at most one page
512 */
513 arghi = 0;
514 BUG_ON(pages > RPCSVC_MAXPAGES);
515 while (pages) {
516 struct page *p = alloc_page(GFP_KERNEL);
517 if (!p)
518 break;
519 rqstp->rq_pages[arghi++] = p;
520 pages--;
521 }
522 return pages == 0;
523}
524
525/*
526 * Release an RPC server buffer
527 */
528static void
529svc_release_buffer(struct svc_rqst *rqstp)
530{
531 unsigned int i;
532
533 for (i = 0; i < ARRAY_SIZE(rqstp->rq_pages); i++)
534 if (rqstp->rq_pages[i])
535 put_page(rqstp->rq_pages[i]);
536}
537
538struct svc_rqst *
539svc_prepare_thread(struct svc_serv *serv, struct svc_pool *pool)
540{
541 struct svc_rqst *rqstp;
542
543 rqstp = kzalloc(sizeof(*rqstp), GFP_KERNEL);
544 if (!rqstp)
545 goto out_enomem;
546
547 init_waitqueue_head(&rqstp->rq_wait);
548
549 serv->sv_nrthreads++;
550 spin_lock_bh(&pool->sp_lock);
551 pool->sp_nrthreads++;
552 list_add(&rqstp->rq_all, &pool->sp_all_threads);
553 spin_unlock_bh(&pool->sp_lock);
554 rqstp->rq_server = serv;
555 rqstp->rq_pool = pool;
556
557 rqstp->rq_argp = kmalloc(serv->sv_xdrsize, GFP_KERNEL);
558 if (!rqstp->rq_argp)
559 goto out_thread;
560
561 rqstp->rq_resp = kmalloc(serv->sv_xdrsize, GFP_KERNEL);
562 if (!rqstp->rq_resp)
563 goto out_thread;
564
565 if (!svc_init_buffer(rqstp, serv->sv_max_mesg))
566 goto out_thread;
567
568 return rqstp;
569out_thread:
570 svc_exit_thread(rqstp);
571out_enomem:
572 return ERR_PTR(-ENOMEM);
573}
574EXPORT_SYMBOL_GPL(svc_prepare_thread);
575
576/*
577 * Choose a pool in which to create a new thread, for svc_set_num_threads
578 */
579static inline struct svc_pool *
580choose_pool(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
581{
582 if (pool != NULL)
583 return pool;
584
585 return &serv->sv_pools[(*state)++ % serv->sv_nrpools];
586}
587
588/*
589 * Choose a thread to kill, for svc_set_num_threads
590 */
591static inline struct task_struct *
592choose_victim(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
593{
594 unsigned int i;
595 struct task_struct *task = NULL;
596
597 if (pool != NULL) {
598 spin_lock_bh(&pool->sp_lock);
599 } else {
600 /* choose a pool in round-robin fashion */
601 for (i = 0; i < serv->sv_nrpools; i++) {
602 pool = &serv->sv_pools[--(*state) % serv->sv_nrpools];
603 spin_lock_bh(&pool->sp_lock);
604 if (!list_empty(&pool->sp_all_threads))
605 goto found_pool;
606 spin_unlock_bh(&pool->sp_lock);
607 }
608 return NULL;
609 }
610
611found_pool:
612 if (!list_empty(&pool->sp_all_threads)) {
613 struct svc_rqst *rqstp;
614
615 /*
616 * Remove from the pool->sp_all_threads list
617 * so we don't try to kill it again.
618 */
619 rqstp = list_entry(pool->sp_all_threads.next, struct svc_rqst, rq_all);
620 list_del_init(&rqstp->rq_all);
621 task = rqstp->rq_task;
622 }
623 spin_unlock_bh(&pool->sp_lock);
624
625 return task;
626}
627
628/*
629 * Create or destroy enough new threads to make the number
630 * of threads the given number. If `pool' is non-NULL, applies
631 * only to threads in that pool, otherwise round-robins between
632 * all pools. Must be called with a svc_get() reference and
633 * the BKL or another lock to protect access to svc_serv fields.
634 *
635 * Destroying threads relies on the service threads filling in
636 * rqstp->rq_task, which only the nfs ones do. Assumes the serv
637 * has been created using svc_create_pooled().
638 *
639 * Based on code that used to be in nfsd_svc() but tweaked
640 * to be pool-aware.
641 */
642int
643svc_set_num_threads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
644{
645 struct svc_rqst *rqstp;
646 struct task_struct *task;
647 struct svc_pool *chosen_pool;
648 int error = 0;
649 unsigned int state = serv->sv_nrthreads-1;
650
651 if (pool == NULL) {
652 /* The -1 assumes caller has done a svc_get() */
653 nrservs -= (serv->sv_nrthreads-1);
654 } else {
655 spin_lock_bh(&pool->sp_lock);
656 nrservs -= pool->sp_nrthreads;
657 spin_unlock_bh(&pool->sp_lock);
658 }
659
660 /* create new threads */
661 while (nrservs > 0) {
662 nrservs--;
663 chosen_pool = choose_pool(serv, pool, &state);
664
665 rqstp = svc_prepare_thread(serv, chosen_pool);
666 if (IS_ERR(rqstp)) {
667 error = PTR_ERR(rqstp);
668 break;
669 }
670
671 __module_get(serv->sv_module);
672 task = kthread_create(serv->sv_function, rqstp, serv->sv_name);
673 if (IS_ERR(task)) {
674 error = PTR_ERR(task);
675 module_put(serv->sv_module);
676 svc_exit_thread(rqstp);
677 break;
678 }
679
680 rqstp->rq_task = task;
681 if (serv->sv_nrpools > 1)
682 svc_pool_map_set_cpumask(task, chosen_pool->sp_id);
683
684 svc_sock_update_bufs(serv);
685 wake_up_process(task);
686 }
687 /* destroy old threads */
688 while (nrservs < 0 &&
689 (task = choose_victim(serv, pool, &state)) != NULL) {
690 send_sig(SIGINT, task, 1);
691 nrservs++;
692 }
693
694 return error;
695}
696EXPORT_SYMBOL_GPL(svc_set_num_threads);
697
698/*
699 * Called from a server thread as it's exiting. Caller must hold the BKL or
700 * the "service mutex", whichever is appropriate for the service.
701 */
702void
703svc_exit_thread(struct svc_rqst *rqstp)
704{
705 struct svc_serv *serv = rqstp->rq_server;
706 struct svc_pool *pool = rqstp->rq_pool;
707
708 svc_release_buffer(rqstp);
709 kfree(rqstp->rq_resp);
710 kfree(rqstp->rq_argp);
711 kfree(rqstp->rq_auth_data);
712
713 spin_lock_bh(&pool->sp_lock);
714 pool->sp_nrthreads--;
715 list_del(&rqstp->rq_all);
716 spin_unlock_bh(&pool->sp_lock);
717
718 kfree(rqstp);
719
720 /* Release the server */
721 if (serv)
722 svc_destroy(serv);
723}
724EXPORT_SYMBOL_GPL(svc_exit_thread);
725
726/*
727 * Register an "inet" protocol family netid with the local
728 * rpcbind daemon via an rpcbind v4 SET request.
729 *
730 * No netconfig infrastructure is available in the kernel, so
731 * we map IP_ protocol numbers to netids by hand.
732 *
733 * Returns zero on success; a negative errno value is returned
734 * if any error occurs.
735 */
736static int __svc_rpcb_register4(const u32 program, const u32 version,
737 const unsigned short protocol,
738 const unsigned short port)
739{
740 const struct sockaddr_in sin = {
741 .sin_family = AF_INET,
742 .sin_addr.s_addr = htonl(INADDR_ANY),
743 .sin_port = htons(port),
744 };
745 const char *netid;
746 int error;
747
748 switch (protocol) {
749 case IPPROTO_UDP:
750 netid = RPCBIND_NETID_UDP;
751 break;
752 case IPPROTO_TCP:
753 netid = RPCBIND_NETID_TCP;
754 break;
755 default:
756 return -ENOPROTOOPT;
757 }
758
759 error = rpcb_v4_register(program, version,
760 (const struct sockaddr *)&sin, netid);
761
762 /*
763 * User space didn't support rpcbind v4, so retry this
764 * registration request with the legacy rpcbind v2 protocol.
765 */
766 if (error == -EPROTONOSUPPORT)
767 error = rpcb_register(program, version, protocol, port);
768
769 return error;
770}
771
772#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
773/*
774 * Register an "inet6" protocol family netid with the local
775 * rpcbind daemon via an rpcbind v4 SET request.
776 *
777 * No netconfig infrastructure is available in the kernel, so
778 * we map IP_ protocol numbers to netids by hand.
779 *
780 * Returns zero on success; a negative errno value is returned
781 * if any error occurs.
782 */
783static int __svc_rpcb_register6(const u32 program, const u32 version,
784 const unsigned short protocol,
785 const unsigned short port)
786{
787 const struct sockaddr_in6 sin6 = {
788 .sin6_family = AF_INET6,
789 .sin6_addr = IN6ADDR_ANY_INIT,
790 .sin6_port = htons(port),
791 };
792 const char *netid;
793 int error;
794
795 switch (protocol) {
796 case IPPROTO_UDP:
797 netid = RPCBIND_NETID_UDP6;
798 break;
799 case IPPROTO_TCP:
800 netid = RPCBIND_NETID_TCP6;
801 break;
802 default:
803 return -ENOPROTOOPT;
804 }
805
806 error = rpcb_v4_register(program, version,
807 (const struct sockaddr *)&sin6, netid);
808
809 /*
810 * User space didn't support rpcbind version 4, so we won't
811 * use a PF_INET6 listener.
812 */
813 if (error == -EPROTONOSUPPORT)
814 error = -EAFNOSUPPORT;
815
816 return error;
817}
818#endif /* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
819
820/*
821 * Register a kernel RPC service via rpcbind version 4.
822 *
823 * Returns zero on success; a negative errno value is returned
824 * if any error occurs.
825 */
826static int __svc_register(const char *progname,
827 const u32 program, const u32 version,
828 const int family,
829 const unsigned short protocol,
830 const unsigned short port)
831{
832 int error = -EAFNOSUPPORT;
833
834 switch (family) {
835 case PF_INET:
836 error = __svc_rpcb_register4(program, version,
837 protocol, port);
838 break;
839#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
840 case PF_INET6:
841 error = __svc_rpcb_register6(program, version,
842 protocol, port);
843#endif /* defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE) */
844 }
845
846 if (error < 0)
847 printk(KERN_WARNING "svc: failed to register %sv%u RPC "
848 "service (errno %d).\n", progname, version, -error);
849 return error;
850}
851
852/**
853 * svc_register - register an RPC service with the local portmapper
854 * @serv: svc_serv struct for the service to register
855 * @family: protocol family of service's listener socket
856 * @proto: transport protocol number to advertise
857 * @port: port to advertise
858 *
859 * Service is registered for any address in the passed-in protocol family
860 */
861int svc_register(const struct svc_serv *serv, const int family,
862 const unsigned short proto, const unsigned short port)
863{
864 struct svc_program *progp;
865 unsigned int i;
866 int error = 0;
867
868 BUG_ON(proto == 0 && port == 0);
869
870 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
871 for (i = 0; i < progp->pg_nvers; i++) {
872 if (progp->pg_vers[i] == NULL)
873 continue;
874
875 dprintk("svc: svc_register(%sv%d, %s, %u, %u)%s\n",
876 progp->pg_name,
877 i,
878 proto == IPPROTO_UDP? "udp" : "tcp",
879 port,
880 family,
881 progp->pg_vers[i]->vs_hidden?
882 " (but not telling portmap)" : "");
883
884 if (progp->pg_vers[i]->vs_hidden)
885 continue;
886
887 error = __svc_register(progp->pg_name, progp->pg_prog,
888 i, family, proto, port);
889 if (error < 0)
890 break;
891 }
892 }
893
894 return error;
895}
896
897/*
898 * If user space is running rpcbind, it should take the v4 UNSET
899 * and clear everything for this [program, version]. If user space
900 * is running portmap, it will reject the v4 UNSET, but won't have
901 * any "inet6" entries anyway. So a PMAP_UNSET should be sufficient
902 * in this case to clear all existing entries for [program, version].
903 */
904static void __svc_unregister(const u32 program, const u32 version,
905 const char *progname)
906{
907 int error;
908
909 error = rpcb_v4_register(program, version, NULL, "");
910
911 /*
912 * User space didn't support rpcbind v4, so retry this
913 * request with the legacy rpcbind v2 protocol.
914 */
915 if (error == -EPROTONOSUPPORT)
916 error = rpcb_register(program, version, 0, 0);
917
918 dprintk("svc: %s(%sv%u), error %d\n",
919 __func__, progname, version, error);
920}
921
922/*
923 * All netids, bind addresses and ports registered for [program, version]
924 * are removed from the local rpcbind database (if the service is not
925 * hidden) to make way for a new instance of the service.
926 *
927 * The result of unregistration is reported via dprintk for those who want
928 * verification of the result, but is otherwise not important.
929 */
930static void svc_unregister(const struct svc_serv *serv)
931{
932 struct svc_program *progp;
933 unsigned long flags;
934 unsigned int i;
935
936 clear_thread_flag(TIF_SIGPENDING);
937
938 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
939 for (i = 0; i < progp->pg_nvers; i++) {
940 if (progp->pg_vers[i] == NULL)
941 continue;
942 if (progp->pg_vers[i]->vs_hidden)
943 continue;
944
945 dprintk("svc: attempting to unregister %sv%u\n",
946 progp->pg_name, i);
947 __svc_unregister(progp->pg_prog, i, progp->pg_name);
948 }
949 }
950
951 spin_lock_irqsave(¤t->sighand->siglock, flags);
952 recalc_sigpending();
953 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
954}
955
956/*
957 * Printk the given error with the address of the client that caused it.
958 */
959static int
960__attribute__ ((format (printf, 2, 3)))
961svc_printk(struct svc_rqst *rqstp, const char *fmt, ...)
962{
963 va_list args;
964 int r;
965 char buf[RPC_MAX_ADDRBUFLEN];
966
967 if (!net_ratelimit())
968 return 0;
969
970 printk(KERN_WARNING "svc: %s: ",
971 svc_print_addr(rqstp, buf, sizeof(buf)));
972
973 va_start(args, fmt);
974 r = vprintk(fmt, args);
975 va_end(args);
976
977 return r;
978}
979
980/*
981 * Common routine for processing the RPC request.
982 */
983static int
984svc_process_common(struct svc_rqst *rqstp, struct kvec *argv, struct kvec *resv)
985{
986 struct svc_program *progp;
987 struct svc_version *versp = NULL; /* compiler food */
988 struct svc_procedure *procp = NULL;
989 struct svc_serv *serv = rqstp->rq_server;
990 kxdrproc_t xdr;
991 __be32 *statp;
992 u32 prog, vers, proc;
993 __be32 auth_stat, rpc_stat;
994 int auth_res;
995 __be32 *reply_statp;
996
997 rpc_stat = rpc_success;
998
999 if (argv->iov_len < 6*4)
1000 goto err_short_len;
1001
1002 /* Will be turned off only in gss privacy case: */
1003 rqstp->rq_splice_ok = 1;
1004 /* Will be turned off only when NFSv4 Sessions are used */
1005 rqstp->rq_usedeferral = 1;
1006 rqstp->rq_dropme = false;
1007
1008 /* Setup reply header */
1009 rqstp->rq_xprt->xpt_ops->xpo_prep_reply_hdr(rqstp);
1010
1011 svc_putu32(resv, rqstp->rq_xid);
1012
1013 vers = svc_getnl(argv);
1014
1015 /* First words of reply: */
1016 svc_putnl(resv, 1); /* REPLY */
1017
1018 if (vers != 2) /* RPC version number */
1019 goto err_bad_rpc;
1020
1021 /* Save position in case we later decide to reject: */
1022 reply_statp = resv->iov_base + resv->iov_len;
1023
1024 svc_putnl(resv, 0); /* ACCEPT */
1025
1026 rqstp->rq_prog = prog = svc_getnl(argv); /* program number */
1027 rqstp->rq_vers = vers = svc_getnl(argv); /* version number */
1028 rqstp->rq_proc = proc = svc_getnl(argv); /* procedure number */
1029
1030 progp = serv->sv_program;
1031
1032 for (progp = serv->sv_program; progp; progp = progp->pg_next)
1033 if (prog == progp->pg_prog)
1034 break;
1035
1036 /*
1037 * Decode auth data, and add verifier to reply buffer.
1038 * We do this before anything else in order to get a decent
1039 * auth verifier.
1040 */
1041 auth_res = svc_authenticate(rqstp, &auth_stat);
1042 /* Also give the program a chance to reject this call: */
1043 if (auth_res == SVC_OK && progp) {
1044 auth_stat = rpc_autherr_badcred;
1045 auth_res = progp->pg_authenticate(rqstp);
1046 }
1047 switch (auth_res) {
1048 case SVC_OK:
1049 break;
1050 case SVC_GARBAGE:
1051 goto err_garbage;
1052 case SVC_SYSERR:
1053 rpc_stat = rpc_system_err;
1054 goto err_bad;
1055 case SVC_DENIED:
1056 goto err_bad_auth;
1057 case SVC_CLOSE:
1058 if (test_bit(XPT_TEMP, &rqstp->rq_xprt->xpt_flags))
1059 svc_close_xprt(rqstp->rq_xprt);
1060 case SVC_DROP:
1061 goto dropit;
1062 case SVC_COMPLETE:
1063 goto sendit;
1064 }
1065
1066 if (progp == NULL)
1067 goto err_bad_prog;
1068
1069 if (vers >= progp->pg_nvers ||
1070 !(versp = progp->pg_vers[vers]))
1071 goto err_bad_vers;
1072
1073 procp = versp->vs_proc + proc;
1074 if (proc >= versp->vs_nproc || !procp->pc_func)
1075 goto err_bad_proc;
1076 rqstp->rq_procinfo = procp;
1077
1078 /* Syntactic check complete */
1079 serv->sv_stats->rpccnt++;
1080
1081 /* Build the reply header. */
1082 statp = resv->iov_base +resv->iov_len;
1083 svc_putnl(resv, RPC_SUCCESS);
1084
1085 /* Bump per-procedure stats counter */
1086 procp->pc_count++;
1087
1088 /* Initialize storage for argp and resp */
1089 memset(rqstp->rq_argp, 0, procp->pc_argsize);
1090 memset(rqstp->rq_resp, 0, procp->pc_ressize);
1091
1092 /* un-reserve some of the out-queue now that we have a
1093 * better idea of reply size
1094 */
1095 if (procp->pc_xdrressize)
1096 svc_reserve_auth(rqstp, procp->pc_xdrressize<<2);
1097
1098 /* Call the function that processes the request. */
1099 if (!versp->vs_dispatch) {
1100 /* Decode arguments */
1101 xdr = procp->pc_decode;
1102 if (xdr && !xdr(rqstp, argv->iov_base, rqstp->rq_argp))
1103 goto err_garbage;
1104
1105 *statp = procp->pc_func(rqstp, rqstp->rq_argp, rqstp->rq_resp);
1106
1107 /* Encode reply */
1108 if (rqstp->rq_dropme) {
1109 if (procp->pc_release)
1110 procp->pc_release(rqstp, NULL, rqstp->rq_resp);
1111 goto dropit;
1112 }
1113 if (*statp == rpc_success &&
1114 (xdr = procp->pc_encode) &&
1115 !xdr(rqstp, resv->iov_base+resv->iov_len, rqstp->rq_resp)) {
1116 dprintk("svc: failed to encode reply\n");
1117 /* serv->sv_stats->rpcsystemerr++; */
1118 *statp = rpc_system_err;
1119 }
1120 } else {
1121 dprintk("svc: calling dispatcher\n");
1122 if (!versp->vs_dispatch(rqstp, statp)) {
1123 /* Release reply info */
1124 if (procp->pc_release)
1125 procp->pc_release(rqstp, NULL, rqstp->rq_resp);
1126 goto dropit;
1127 }
1128 }
1129
1130 /* Check RPC status result */
1131 if (*statp != rpc_success)
1132 resv->iov_len = ((void*)statp) - resv->iov_base + 4;
1133
1134 /* Release reply info */
1135 if (procp->pc_release)
1136 procp->pc_release(rqstp, NULL, rqstp->rq_resp);
1137
1138 if (procp->pc_encode == NULL)
1139 goto dropit;
1140
1141 sendit:
1142 if (svc_authorise(rqstp))
1143 goto dropit;
1144 return 1; /* Caller can now send it */
1145
1146 dropit:
1147 svc_authorise(rqstp); /* doesn't hurt to call this twice */
1148 dprintk("svc: svc_process dropit\n");
1149 return 0;
1150
1151err_short_len:
1152 svc_printk(rqstp, "short len %Zd, dropping request\n",
1153 argv->iov_len);
1154
1155 goto dropit; /* drop request */
1156
1157err_bad_rpc:
1158 serv->sv_stats->rpcbadfmt++;
1159 svc_putnl(resv, 1); /* REJECT */
1160 svc_putnl(resv, 0); /* RPC_MISMATCH */
1161 svc_putnl(resv, 2); /* Only RPCv2 supported */
1162 svc_putnl(resv, 2);
1163 goto sendit;
1164
1165err_bad_auth:
1166 dprintk("svc: authentication failed (%d)\n", ntohl(auth_stat));
1167 serv->sv_stats->rpcbadauth++;
1168 /* Restore write pointer to location of accept status: */
1169 xdr_ressize_check(rqstp, reply_statp);
1170 svc_putnl(resv, 1); /* REJECT */
1171 svc_putnl(resv, 1); /* AUTH_ERROR */
1172 svc_putnl(resv, ntohl(auth_stat)); /* status */
1173 goto sendit;
1174
1175err_bad_prog:
1176 dprintk("svc: unknown program %d\n", prog);
1177 serv->sv_stats->rpcbadfmt++;
1178 svc_putnl(resv, RPC_PROG_UNAVAIL);
1179 goto sendit;
1180
1181err_bad_vers:
1182 svc_printk(rqstp, "unknown version (%d for prog %d, %s)\n",
1183 vers, prog, progp->pg_name);
1184
1185 serv->sv_stats->rpcbadfmt++;
1186 svc_putnl(resv, RPC_PROG_MISMATCH);
1187 svc_putnl(resv, progp->pg_lovers);
1188 svc_putnl(resv, progp->pg_hivers);
1189 goto sendit;
1190
1191err_bad_proc:
1192 svc_printk(rqstp, "unknown procedure (%d)\n", proc);
1193
1194 serv->sv_stats->rpcbadfmt++;
1195 svc_putnl(resv, RPC_PROC_UNAVAIL);
1196 goto sendit;
1197
1198err_garbage:
1199 svc_printk(rqstp, "failed to decode args\n");
1200
1201 rpc_stat = rpc_garbage_args;
1202err_bad:
1203 serv->sv_stats->rpcbadfmt++;
1204 svc_putnl(resv, ntohl(rpc_stat));
1205 goto sendit;
1206}
1207EXPORT_SYMBOL_GPL(svc_process);
1208
1209/*
1210 * Process the RPC request.
1211 */
1212int
1213svc_process(struct svc_rqst *rqstp)
1214{
1215 struct kvec *argv = &rqstp->rq_arg.head[0];
1216 struct kvec *resv = &rqstp->rq_res.head[0];
1217 struct svc_serv *serv = rqstp->rq_server;
1218 u32 dir;
1219
1220 /*
1221 * Setup response xdr_buf.
1222 * Initially it has just one page
1223 */
1224 rqstp->rq_resused = 1;
1225 resv->iov_base = page_address(rqstp->rq_respages[0]);
1226 resv->iov_len = 0;
1227 rqstp->rq_res.pages = rqstp->rq_respages + 1;
1228 rqstp->rq_res.len = 0;
1229 rqstp->rq_res.page_base = 0;
1230 rqstp->rq_res.page_len = 0;
1231 rqstp->rq_res.buflen = PAGE_SIZE;
1232 rqstp->rq_res.tail[0].iov_base = NULL;
1233 rqstp->rq_res.tail[0].iov_len = 0;
1234
1235 rqstp->rq_xid = svc_getu32(argv);
1236
1237 dir = svc_getnl(argv);
1238 if (dir != 0) {
1239 /* direction != CALL */
1240 svc_printk(rqstp, "bad direction %d, dropping request\n", dir);
1241 serv->sv_stats->rpcbadfmt++;
1242 svc_drop(rqstp);
1243 return 0;
1244 }
1245
1246 /* Returns 1 for send, 0 for drop */
1247 if (svc_process_common(rqstp, argv, resv))
1248 return svc_send(rqstp);
1249 else {
1250 svc_drop(rqstp);
1251 return 0;
1252 }
1253}
1254
1255#if defined(CONFIG_SUNRPC_BACKCHANNEL)
1256/*
1257 * Process a backchannel RPC request that arrived over an existing
1258 * outbound connection
1259 */
1260int
1261bc_svc_process(struct svc_serv *serv, struct rpc_rqst *req,
1262 struct svc_rqst *rqstp)
1263{
1264 struct kvec *argv = &rqstp->rq_arg.head[0];
1265 struct kvec *resv = &rqstp->rq_res.head[0];
1266
1267 /* Build the svc_rqst used by the common processing routine */
1268 rqstp->rq_xprt = serv->sv_bc_xprt;
1269 rqstp->rq_xid = req->rq_xid;
1270 rqstp->rq_prot = req->rq_xprt->prot;
1271 rqstp->rq_server = serv;
1272
1273 rqstp->rq_addrlen = sizeof(req->rq_xprt->addr);
1274 memcpy(&rqstp->rq_addr, &req->rq_xprt->addr, rqstp->rq_addrlen);
1275 memcpy(&rqstp->rq_arg, &req->rq_rcv_buf, sizeof(rqstp->rq_arg));
1276 memcpy(&rqstp->rq_res, &req->rq_snd_buf, sizeof(rqstp->rq_res));
1277
1278 /* reset result send buffer "put" position */
1279 resv->iov_len = 0;
1280
1281 if (rqstp->rq_prot != IPPROTO_TCP) {
1282 printk(KERN_ERR "No support for Non-TCP transports!\n");
1283 BUG();
1284 }
1285
1286 /*
1287 * Skip the next two words because they've already been
1288 * processed in the trasport
1289 */
1290 svc_getu32(argv); /* XID */
1291 svc_getnl(argv); /* CALLDIR */
1292
1293 /* Returns 1 for send, 0 for drop */
1294 if (svc_process_common(rqstp, argv, resv)) {
1295 memcpy(&req->rq_snd_buf, &rqstp->rq_res,
1296 sizeof(req->rq_snd_buf));
1297 return bc_send(req);
1298 } else {
1299 /* Nothing to do to drop request */
1300 return 0;
1301 }
1302}
1303EXPORT_SYMBOL_GPL(bc_svc_process);
1304#endif /* CONFIG_SUNRPC_BACKCHANNEL */
1305
1306/*
1307 * Return (transport-specific) limit on the rpc payload.
1308 */
1309u32 svc_max_payload(const struct svc_rqst *rqstp)
1310{
1311 u32 max = rqstp->rq_xprt->xpt_class->xcl_max_payload;
1312
1313 if (rqstp->rq_server->sv_max_payload < max)
1314 max = rqstp->rq_server->sv_max_payload;
1315 return max;
1316}
1317EXPORT_SYMBOL_GPL(svc_max_payload);