2 * linux/net/sunrpc/svc.c
4 * High-level RPC service routines
6 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
8 * Multiple threads pools and NUMAisation
9 * Copyright (c) 2006 Silicon Graphics, Inc.
10 * by Greg Banks <gnb@melbourne.sgi.com>
13 #include <linux/linkage.h>
14 #include <linux/sched.h>
15 #include <linux/errno.h>
16 #include <linux/net.h>
19 #include <linux/interrupt.h>
20 #include <linux/module.h>
22 #include <linux/sunrpc/types.h>
23 #include <linux/sunrpc/xdr.h>
24 #include <linux/sunrpc/stats.h>
25 #include <linux/sunrpc/svcsock.h>
26 #include <linux/sunrpc/clnt.h>
28 #define RPCDBG_FACILITY RPCDBG_SVCDSP
29 #define RPC_PARANOIA 1
32 * Mode for mapping cpus to pools.
35 SVC_POOL_NONE = -1, /* uninitialised, choose one of the others */
36 SVC_POOL_GLOBAL, /* no mapping, just a single global pool
37 * (legacy & UP mode) */
38 SVC_POOL_PERCPU, /* one pool per cpu */
39 SVC_POOL_PERNODE /* one pool per numa node */
43 * Structure for mapping cpus to pools and vice versa.
44 * Setup once during sunrpc initialisation.
46 static struct svc_pool_map {
47 int mode; /* Note: int not enum to avoid
48 * warnings about "enumeration value
49 * not handled in switch" */
51 unsigned int *pool_to; /* maps pool id to cpu or node */
52 unsigned int *to_pool; /* maps cpu or node to pool id */
59 * Detect best pool mapping mode heuristically,
60 * according to the machine's topology.
63 svc_pool_map_choose_mode(void)
67 if (num_online_nodes() > 1) {
69 * Actually have multiple NUMA nodes,
70 * so split pools on NUMA node boundaries
72 return SVC_POOL_PERNODE;
75 node = any_online_node(node_online_map);
76 if (nr_cpus_node(node) > 2) {
78 * Non-trivial SMP, or CONFIG_NUMA on
79 * non-NUMA hardware, e.g. with a generic
80 * x86_64 kernel on Xeons. In this case we
81 * want to divide the pools on cpu boundaries.
83 return SVC_POOL_PERCPU;
86 /* default: one global pool */
87 return SVC_POOL_GLOBAL;
91 * Allocate the to_pool[] and pool_to[] arrays.
92 * Returns 0 on success or an errno.
95 svc_pool_map_alloc_arrays(struct svc_pool_map *m, unsigned int maxpools)
97 m->to_pool = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
100 m->pool_to = kcalloc(maxpools, sizeof(unsigned int), GFP_KERNEL);
113 * Initialise the pool map for SVC_POOL_PERCPU mode.
114 * Returns number of pools or <0 on error.
117 svc_pool_map_init_percpu(struct svc_pool_map *m)
119 unsigned int maxpools = highest_possible_processor_id()+1;
120 unsigned int pidx = 0;
124 err = svc_pool_map_alloc_arrays(m, maxpools);
128 for_each_online_cpu(cpu) {
129 BUG_ON(pidx > maxpools);
130 m->to_pool[cpu] = pidx;
131 m->pool_to[pidx] = cpu;
134 /* cpus brought online later all get mapped to pool0, sorry */
141 * Initialise the pool map for SVC_POOL_PERNODE mode.
142 * Returns number of pools or <0 on error.
145 svc_pool_map_init_pernode(struct svc_pool_map *m)
147 unsigned int maxpools = highest_possible_node_id()+1;
148 unsigned int pidx = 0;
152 err = svc_pool_map_alloc_arrays(m, maxpools);
156 for_each_node_with_cpus(node) {
157 /* some architectures (e.g. SN2) have cpuless nodes */
158 BUG_ON(pidx > maxpools);
159 m->to_pool[node] = pidx;
160 m->pool_to[pidx] = node;
163 /* nodes brought online later all get mapped to pool0, sorry */
170 * Build the global map of cpus to pools and vice versa.
173 svc_pool_map_init(void)
175 struct svc_pool_map *m = &svc_pool_map;
178 if (m->mode != SVC_POOL_NONE)
181 m->mode = svc_pool_map_choose_mode();
184 case SVC_POOL_PERCPU:
185 npools = svc_pool_map_init_percpu(m);
187 case SVC_POOL_PERNODE:
188 npools = svc_pool_map_init_pernode(m);
193 /* default, or memory allocation failure */
195 m->mode = SVC_POOL_GLOBAL;
203 * Set the current thread's cpus_allowed mask so that it
204 * will only run on cpus in the given pool.
206 * Returns 1 and fills in oldmask iff a cpumask was applied.
209 svc_pool_map_set_cpumask(unsigned int pidx, cpumask_t *oldmask)
211 struct svc_pool_map *m = &svc_pool_map;
212 unsigned int node; /* or cpu */
215 * The caller checks for sv_nrpools > 1, which
216 * implies that we've been initialized and the
217 * map mode is not NONE.
219 BUG_ON(m->mode == SVC_POOL_NONE);
225 case SVC_POOL_PERCPU:
226 node = m->pool_to[pidx];
227 *oldmask = current->cpus_allowed;
228 set_cpus_allowed(current, cpumask_of_cpu(node));
230 case SVC_POOL_PERNODE:
231 node = m->pool_to[pidx];
232 *oldmask = current->cpus_allowed;
233 set_cpus_allowed(current, node_to_cpumask(node));
239 * Use the mapping mode to choose a pool for a given CPU.
240 * Used when enqueueing an incoming RPC. Always returns
241 * a non-NULL pool pointer.
244 svc_pool_for_cpu(struct svc_serv *serv, int cpu)
246 struct svc_pool_map *m = &svc_pool_map;
247 unsigned int pidx = 0;
250 * SVC_POOL_NONE happens in a pure client when
251 * lockd is brought up, so silently treat it the
252 * same as SVC_POOL_GLOBAL.
256 case SVC_POOL_PERCPU:
257 pidx = m->to_pool[cpu];
259 case SVC_POOL_PERNODE:
260 pidx = m->to_pool[cpu_to_node(cpu)];
263 return &serv->sv_pools[pidx % serv->sv_nrpools];
268 * Create an RPC service
270 static struct svc_serv *
271 __svc_create(struct svc_program *prog, unsigned int bufsize, int npools,
272 void (*shutdown)(struct svc_serv *serv))
274 struct svc_serv *serv;
276 unsigned int xdrsize;
279 if (!(serv = kzalloc(sizeof(*serv), GFP_KERNEL)))
281 serv->sv_name = prog->pg_name;
282 serv->sv_program = prog;
283 serv->sv_nrthreads = 1;
284 serv->sv_stats = prog->pg_stats;
285 if (bufsize > RPCSVC_MAXPAYLOAD)
286 bufsize = RPCSVC_MAXPAYLOAD;
287 serv->sv_max_payload = bufsize? bufsize : 4096;
288 serv->sv_max_mesg = roundup(serv->sv_max_payload + PAGE_SIZE, PAGE_SIZE);
289 serv->sv_shutdown = shutdown;
292 prog->pg_lovers = prog->pg_nvers-1;
293 for (vers=0; vers<prog->pg_nvers ; vers++)
294 if (prog->pg_vers[vers]) {
295 prog->pg_hivers = vers;
296 if (prog->pg_lovers > vers)
297 prog->pg_lovers = vers;
298 if (prog->pg_vers[vers]->vs_xdrsize > xdrsize)
299 xdrsize = prog->pg_vers[vers]->vs_xdrsize;
301 prog = prog->pg_next;
303 serv->sv_xdrsize = xdrsize;
304 INIT_LIST_HEAD(&serv->sv_tempsocks);
305 INIT_LIST_HEAD(&serv->sv_permsocks);
306 init_timer(&serv->sv_temptimer);
307 spin_lock_init(&serv->sv_lock);
309 serv->sv_nrpools = npools;
311 kcalloc(sizeof(struct svc_pool), serv->sv_nrpools,
313 if (!serv->sv_pools) {
318 for (i = 0; i < serv->sv_nrpools; i++) {
319 struct svc_pool *pool = &serv->sv_pools[i];
321 dprintk("initialising pool %u for %s\n",
325 INIT_LIST_HEAD(&pool->sp_threads);
326 INIT_LIST_HEAD(&pool->sp_sockets);
327 INIT_LIST_HEAD(&pool->sp_all_threads);
328 spin_lock_init(&pool->sp_lock);
332 /* Remove any stale portmap registrations */
333 svc_register(serv, 0, 0);
339 svc_create(struct svc_program *prog, unsigned int bufsize,
340 void (*shutdown)(struct svc_serv *serv))
342 return __svc_create(prog, bufsize, /*npools*/1, shutdown);
346 svc_create_pooled(struct svc_program *prog, unsigned int bufsize,
347 void (*shutdown)(struct svc_serv *serv),
348 svc_thread_fn func, int sig, struct module *mod)
350 struct svc_serv *serv;
351 unsigned int npools = svc_pool_map_init();
353 serv = __svc_create(prog, bufsize, npools, shutdown);
356 serv->sv_function = func;
357 serv->sv_kill_signal = sig;
358 serv->sv_module = mod;
365 * Destroy an RPC service. Should be called with the BKL held
368 svc_destroy(struct svc_serv *serv)
370 struct svc_sock *svsk;
372 dprintk("RPC: svc_destroy(%s, %d)\n",
373 serv->sv_program->pg_name,
376 if (serv->sv_nrthreads) {
377 if (--(serv->sv_nrthreads) != 0) {
378 svc_sock_update_bufs(serv);
382 printk("svc_destroy: no threads for serv=%p!\n", serv);
384 del_timer_sync(&serv->sv_temptimer);
386 while (!list_empty(&serv->sv_tempsocks)) {
387 svsk = list_entry(serv->sv_tempsocks.next,
390 svc_delete_socket(svsk);
392 if (serv->sv_shutdown)
393 serv->sv_shutdown(serv);
395 while (!list_empty(&serv->sv_permsocks)) {
396 svsk = list_entry(serv->sv_permsocks.next,
399 svc_delete_socket(svsk);
402 cache_clean_deferred(serv);
404 /* Unregister service with the portmapper */
405 svc_register(serv, 0, 0);
406 kfree(serv->sv_pools);
411 * Allocate an RPC server's buffer space.
412 * We allocate pages and place them in rq_argpages.
415 svc_init_buffer(struct svc_rqst *rqstp, unsigned int size)
420 pages = size / PAGE_SIZE + 1; /* extra page as we hold both request and reply.
421 * We assume one is at most one page
424 BUG_ON(pages > RPCSVC_MAXPAGES);
426 struct page *p = alloc_page(GFP_KERNEL);
429 rqstp->rq_pages[arghi++] = p;
436 * Release an RPC server buffer
439 svc_release_buffer(struct svc_rqst *rqstp)
442 for (i=0; i<ARRAY_SIZE(rqstp->rq_pages); i++)
443 if (rqstp->rq_pages[i])
444 put_page(rqstp->rq_pages[i]);
448 * Create a thread in the given pool. Caller must hold BKL.
449 * On a NUMA or SMP machine, with a multi-pool serv, the thread
450 * will be restricted to run on the cpus belonging to the pool.
453 __svc_create_thread(svc_thread_fn func, struct svc_serv *serv,
454 struct svc_pool *pool)
456 struct svc_rqst *rqstp;
458 int have_oldmask = 0;
461 rqstp = kzalloc(sizeof(*rqstp), GFP_KERNEL);
465 init_waitqueue_head(&rqstp->rq_wait);
467 if (!(rqstp->rq_argp = kmalloc(serv->sv_xdrsize, GFP_KERNEL))
468 || !(rqstp->rq_resp = kmalloc(serv->sv_xdrsize, GFP_KERNEL))
469 || !svc_init_buffer(rqstp, serv->sv_max_mesg))
472 serv->sv_nrthreads++;
473 spin_lock_bh(&pool->sp_lock);
474 pool->sp_nrthreads++;
475 list_add(&rqstp->rq_all, &pool->sp_all_threads);
476 spin_unlock_bh(&pool->sp_lock);
477 rqstp->rq_server = serv;
478 rqstp->rq_pool = pool;
480 if (serv->sv_nrpools > 1)
481 have_oldmask = svc_pool_map_set_cpumask(pool->sp_id, &oldmask);
483 error = kernel_thread((int (*)(void *)) func, rqstp, 0);
486 set_cpus_allowed(current, oldmask);
490 svc_sock_update_bufs(serv);
496 svc_exit_thread(rqstp);
501 * Create a thread in the default pool. Caller must hold BKL.
504 svc_create_thread(svc_thread_fn func, struct svc_serv *serv)
506 return __svc_create_thread(func, serv, &serv->sv_pools[0]);
510 * Choose a pool in which to create a new thread, for svc_set_num_threads
512 static inline struct svc_pool *
513 choose_pool(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
518 return &serv->sv_pools[(*state)++ % serv->sv_nrpools];
522 * Choose a thread to kill, for svc_set_num_threads
524 static inline struct task_struct *
525 choose_victim(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
528 struct task_struct *task = NULL;
531 spin_lock_bh(&pool->sp_lock);
533 /* choose a pool in round-robin fashion */
534 for (i = 0; i < serv->sv_nrpools; i++) {
535 pool = &serv->sv_pools[--(*state) % serv->sv_nrpools];
536 spin_lock_bh(&pool->sp_lock);
537 if (!list_empty(&pool->sp_all_threads))
539 spin_unlock_bh(&pool->sp_lock);
545 if (!list_empty(&pool->sp_all_threads)) {
546 struct svc_rqst *rqstp;
549 * Remove from the pool->sp_all_threads list
550 * so we don't try to kill it again.
552 rqstp = list_entry(pool->sp_all_threads.next, struct svc_rqst, rq_all);
553 list_del_init(&rqstp->rq_all);
554 task = rqstp->rq_task;
556 spin_unlock_bh(&pool->sp_lock);
562 * Create or destroy enough new threads to make the number
563 * of threads the given number. If `pool' is non-NULL, applies
564 * only to threads in that pool, otherwise round-robins between
565 * all pools. Must be called with a svc_get() reference and
568 * Destroying threads relies on the service threads filling in
569 * rqstp->rq_task, which only the nfs ones do. Assumes the serv
570 * has been created using svc_create_pooled().
572 * Based on code that used to be in nfsd_svc() but tweaked
576 svc_set_num_threads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
578 struct task_struct *victim;
580 unsigned int state = serv->sv_nrthreads-1;
583 /* The -1 assumes caller has done a svc_get() */
584 nrservs -= (serv->sv_nrthreads-1);
586 spin_lock_bh(&pool->sp_lock);
587 nrservs -= pool->sp_nrthreads;
588 spin_unlock_bh(&pool->sp_lock);
591 /* create new threads */
592 while (nrservs > 0) {
594 __module_get(serv->sv_module);
595 error = __svc_create_thread(serv->sv_function, serv,
596 choose_pool(serv, pool, &state));
598 module_put(serv->sv_module);
602 /* destroy old threads */
603 while (nrservs < 0 &&
604 (victim = choose_victim(serv, pool, &state)) != NULL) {
605 send_sig(serv->sv_kill_signal, victim, 1);
613 * Called from a server thread as it's exiting. Caller must hold BKL.
616 svc_exit_thread(struct svc_rqst *rqstp)
618 struct svc_serv *serv = rqstp->rq_server;
619 struct svc_pool *pool = rqstp->rq_pool;
621 svc_release_buffer(rqstp);
622 kfree(rqstp->rq_resp);
623 kfree(rqstp->rq_argp);
624 kfree(rqstp->rq_auth_data);
626 spin_lock_bh(&pool->sp_lock);
627 pool->sp_nrthreads--;
628 list_del(&rqstp->rq_all);
629 spin_unlock_bh(&pool->sp_lock);
633 /* Release the server */
639 * Register an RPC service with the local portmapper.
640 * To unregister a service, call this routine with
641 * proto and port == 0.
644 svc_register(struct svc_serv *serv, int proto, unsigned short port)
646 struct svc_program *progp;
648 int i, error = 0, dummy;
651 clear_thread_flag(TIF_SIGPENDING);
653 for (progp = serv->sv_program; progp; progp = progp->pg_next) {
654 for (i = 0; i < progp->pg_nvers; i++) {
655 if (progp->pg_vers[i] == NULL)
658 dprintk("RPC: svc_register(%s, %s, %d, %d)%s\n",
660 proto == IPPROTO_UDP? "udp" : "tcp",
663 progp->pg_vers[i]->vs_hidden?
664 " (but not telling portmap)" : "");
666 if (progp->pg_vers[i]->vs_hidden)
669 error = rpc_register(progp->pg_prog, i, proto, port, &dummy);
672 if (port && !dummy) {
680 spin_lock_irqsave(¤t->sighand->siglock, flags);
682 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
689 * Process the RPC request.
692 svc_process(struct svc_rqst *rqstp)
694 struct svc_program *progp;
695 struct svc_version *versp = NULL; /* compiler food */
696 struct svc_procedure *procp = NULL;
697 struct kvec * argv = &rqstp->rq_arg.head[0];
698 struct kvec * resv = &rqstp->rq_res.head[0];
699 struct svc_serv *serv = rqstp->rq_server;
702 u32 dir, prog, vers, proc;
703 __be32 auth_stat, rpc_stat;
707 rpc_stat = rpc_success;
709 if (argv->iov_len < 6*4)
712 /* setup response xdr_buf.
713 * Initially it has just one page
715 rqstp->rq_resused = 1;
716 resv->iov_base = page_address(rqstp->rq_respages[0]);
718 rqstp->rq_res.pages = rqstp->rq_respages + 1;
719 rqstp->rq_res.len = 0;
720 rqstp->rq_res.page_base = 0;
721 rqstp->rq_res.page_len = 0;
722 rqstp->rq_res.buflen = PAGE_SIZE;
723 rqstp->rq_res.tail[0].iov_base = NULL;
724 rqstp->rq_res.tail[0].iov_len = 0;
725 /* Will be turned off only in gss privacy case: */
726 rqstp->rq_sendfile_ok = 1;
727 /* tcp needs a space for the record length... */
728 if (rqstp->rq_prot == IPPROTO_TCP)
731 rqstp->rq_xid = svc_getu32(argv);
732 svc_putu32(resv, rqstp->rq_xid);
734 dir = svc_getnl(argv);
735 vers = svc_getnl(argv);
737 /* First words of reply: */
738 svc_putnl(resv, 1); /* REPLY */
740 if (dir != 0) /* direction != CALL */
742 if (vers != 2) /* RPC version number */
745 /* Save position in case we later decide to reject: */
746 reply_statp = resv->iov_base + resv->iov_len;
748 svc_putnl(resv, 0); /* ACCEPT */
750 rqstp->rq_prog = prog = svc_getnl(argv); /* program number */
751 rqstp->rq_vers = vers = svc_getnl(argv); /* version number */
752 rqstp->rq_proc = proc = svc_getnl(argv); /* procedure number */
754 progp = serv->sv_program;
756 for (progp = serv->sv_program; progp; progp = progp->pg_next)
757 if (prog == progp->pg_prog)
761 * Decode auth data, and add verifier to reply buffer.
762 * We do this before anything else in order to get a decent
765 auth_res = svc_authenticate(rqstp, &auth_stat);
766 /* Also give the program a chance to reject this call: */
767 if (auth_res == SVC_OK && progp) {
768 auth_stat = rpc_autherr_badcred;
769 auth_res = progp->pg_authenticate(rqstp);
775 rpc_stat = rpc_garbage_args;
778 rpc_stat = rpc_system_err;
791 if (vers >= progp->pg_nvers ||
792 !(versp = progp->pg_vers[vers]))
795 procp = versp->vs_proc + proc;
796 if (proc >= versp->vs_nproc || !procp->pc_func)
798 rqstp->rq_server = serv;
799 rqstp->rq_procinfo = procp;
801 /* Syntactic check complete */
802 serv->sv_stats->rpccnt++;
804 /* Build the reply header. */
805 statp = resv->iov_base +resv->iov_len;
806 svc_putnl(resv, RPC_SUCCESS);
808 /* Bump per-procedure stats counter */
811 /* Initialize storage for argp and resp */
812 memset(rqstp->rq_argp, 0, procp->pc_argsize);
813 memset(rqstp->rq_resp, 0, procp->pc_ressize);
815 /* un-reserve some of the out-queue now that we have a
816 * better idea of reply size
818 if (procp->pc_xdrressize)
819 svc_reserve(rqstp, procp->pc_xdrressize<<2);
821 /* Call the function that processes the request. */
822 if (!versp->vs_dispatch) {
823 /* Decode arguments */
824 xdr = procp->pc_decode;
825 if (xdr && !xdr(rqstp, argv->iov_base, rqstp->rq_argp))
828 *statp = procp->pc_func(rqstp, rqstp->rq_argp, rqstp->rq_resp);
831 if (*statp == rpc_success && (xdr = procp->pc_encode)
832 && !xdr(rqstp, resv->iov_base+resv->iov_len, rqstp->rq_resp)) {
833 dprintk("svc: failed to encode reply\n");
834 /* serv->sv_stats->rpcsystemerr++; */
835 *statp = rpc_system_err;
838 dprintk("svc: calling dispatcher\n");
839 if (!versp->vs_dispatch(rqstp, statp)) {
840 /* Release reply info */
841 if (procp->pc_release)
842 procp->pc_release(rqstp, NULL, rqstp->rq_resp);
847 /* Check RPC status result */
848 if (*statp != rpc_success)
849 resv->iov_len = ((void*)statp) - resv->iov_base + 4;
851 /* Release reply info */
852 if (procp->pc_release)
853 procp->pc_release(rqstp, NULL, rqstp->rq_resp);
855 if (procp->pc_encode == NULL)
859 if (svc_authorise(rqstp))
861 return svc_send(rqstp);
864 svc_authorise(rqstp); /* doesn't hurt to call this twice */
865 dprintk("svc: svc_process dropit\n");
871 printk("svc: short len %Zd, dropping request\n", argv->iov_len);
873 goto dropit; /* drop request */
877 printk("svc: bad direction %d, dropping request\n", dir);
879 serv->sv_stats->rpcbadfmt++;
880 goto dropit; /* drop request */
883 serv->sv_stats->rpcbadfmt++;
884 svc_putnl(resv, 1); /* REJECT */
885 svc_putnl(resv, 0); /* RPC_MISMATCH */
886 svc_putnl(resv, 2); /* Only RPCv2 supported */
891 dprintk("svc: authentication failed (%d)\n", ntohl(auth_stat));
892 serv->sv_stats->rpcbadauth++;
893 /* Restore write pointer to location of accept status: */
894 xdr_ressize_check(rqstp, reply_statp);
895 svc_putnl(resv, 1); /* REJECT */
896 svc_putnl(resv, 1); /* AUTH_ERROR */
897 svc_putnl(resv, ntohl(auth_stat)); /* status */
901 dprintk("svc: unknown program %d\n", prog);
902 serv->sv_stats->rpcbadfmt++;
903 svc_putnl(resv, RPC_PROG_UNAVAIL);
908 printk("svc: unknown version (%d)\n", vers);
910 serv->sv_stats->rpcbadfmt++;
911 svc_putnl(resv, RPC_PROG_MISMATCH);
912 svc_putnl(resv, progp->pg_lovers);
913 svc_putnl(resv, progp->pg_hivers);
918 printk("svc: unknown procedure (%d)\n", proc);
920 serv->sv_stats->rpcbadfmt++;
921 svc_putnl(resv, RPC_PROC_UNAVAIL);
926 printk("svc: failed to decode args\n");
928 rpc_stat = rpc_garbage_args;
930 serv->sv_stats->rpcbadfmt++;
931 svc_putnl(resv, ntohl(rpc_stat));
936 * Return (transport-specific) limit on the rpc payload.
938 u32 svc_max_payload(const struct svc_rqst *rqstp)
940 int max = RPCSVC_MAXPAYLOAD_TCP;
942 if (rqstp->rq_sock->sk_sock->type == SOCK_DGRAM)
943 max = RPCSVC_MAXPAYLOAD_UDP;
944 if (rqstp->rq_server->sv_max_payload < max)
945 max = rqstp->rq_server->sv_max_payload;
948 EXPORT_SYMBOL_GPL(svc_max_payload);