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 serv->sv_bufsz = bufsize? bufsize : 4096;
286 serv->sv_shutdown = shutdown;
289 prog->pg_lovers = prog->pg_nvers-1;
290 for (vers=0; vers<prog->pg_nvers ; vers++)
291 if (prog->pg_vers[vers]) {
292 prog->pg_hivers = vers;
293 if (prog->pg_lovers > vers)
294 prog->pg_lovers = vers;
295 if (prog->pg_vers[vers]->vs_xdrsize > xdrsize)
296 xdrsize = prog->pg_vers[vers]->vs_xdrsize;
298 prog = prog->pg_next;
300 serv->sv_xdrsize = xdrsize;
301 INIT_LIST_HEAD(&serv->sv_tempsocks);
302 INIT_LIST_HEAD(&serv->sv_permsocks);
303 init_timer(&serv->sv_temptimer);
304 spin_lock_init(&serv->sv_lock);
306 serv->sv_nrpools = npools;
308 kcalloc(sizeof(struct svc_pool), serv->sv_nrpools,
310 if (!serv->sv_pools) {
315 for (i = 0; i < serv->sv_nrpools; i++) {
316 struct svc_pool *pool = &serv->sv_pools[i];
318 dprintk("initialising pool %u for %s\n",
322 INIT_LIST_HEAD(&pool->sp_threads);
323 INIT_LIST_HEAD(&pool->sp_sockets);
324 INIT_LIST_HEAD(&pool->sp_all_threads);
325 spin_lock_init(&pool->sp_lock);
329 /* Remove any stale portmap registrations */
330 svc_register(serv, 0, 0);
336 svc_create(struct svc_program *prog, unsigned int bufsize,
337 void (*shutdown)(struct svc_serv *serv))
339 return __svc_create(prog, bufsize, /*npools*/1, shutdown);
343 svc_create_pooled(struct svc_program *prog, unsigned int bufsize,
344 void (*shutdown)(struct svc_serv *serv),
345 svc_thread_fn func, int sig, struct module *mod)
347 struct svc_serv *serv;
348 unsigned int npools = svc_pool_map_init();
350 serv = __svc_create(prog, bufsize, npools, shutdown);
353 serv->sv_function = func;
354 serv->sv_kill_signal = sig;
355 serv->sv_module = mod;
362 * Destroy an RPC service. Should be called with the BKL held
365 svc_destroy(struct svc_serv *serv)
367 struct svc_sock *svsk;
369 dprintk("RPC: svc_destroy(%s, %d)\n",
370 serv->sv_program->pg_name,
373 if (serv->sv_nrthreads) {
374 if (--(serv->sv_nrthreads) != 0) {
375 svc_sock_update_bufs(serv);
379 printk("svc_destroy: no threads for serv=%p!\n", serv);
381 del_timer_sync(&serv->sv_temptimer);
383 while (!list_empty(&serv->sv_tempsocks)) {
384 svsk = list_entry(serv->sv_tempsocks.next,
387 svc_delete_socket(svsk);
389 if (serv->sv_shutdown)
390 serv->sv_shutdown(serv);
392 while (!list_empty(&serv->sv_permsocks)) {
393 svsk = list_entry(serv->sv_permsocks.next,
396 svc_delete_socket(svsk);
399 cache_clean_deferred(serv);
401 /* Unregister service with the portmapper */
402 svc_register(serv, 0, 0);
403 kfree(serv->sv_pools);
408 * Allocate an RPC server's buffer space.
409 * We allocate pages and place them in rq_argpages.
412 svc_init_buffer(struct svc_rqst *rqstp, unsigned int size)
417 if (size > RPCSVC_MAXPAYLOAD)
418 size = RPCSVC_MAXPAYLOAD;
419 pages = 2 + (size+ PAGE_SIZE -1) / PAGE_SIZE;
420 rqstp->rq_argused = 0;
421 rqstp->rq_resused = 0;
423 BUG_ON(pages > RPCSVC_MAXPAGES);
425 struct page *p = alloc_page(GFP_KERNEL);
428 rqstp->rq_argpages[arghi++] = p;
431 rqstp->rq_arghi = arghi;
436 * Release an RPC server buffer
439 svc_release_buffer(struct svc_rqst *rqstp)
441 while (rqstp->rq_arghi)
442 put_page(rqstp->rq_argpages[--rqstp->rq_arghi]);
443 while (rqstp->rq_resused) {
444 if (rqstp->rq_respages[--rqstp->rq_resused] == NULL)
446 put_page(rqstp->rq_respages[rqstp->rq_resused]);
448 rqstp->rq_argused = 0;
452 * Create a thread in the given pool. Caller must hold BKL.
453 * On a NUMA or SMP machine, with a multi-pool serv, the thread
454 * will be restricted to run on the cpus belonging to the pool.
457 __svc_create_thread(svc_thread_fn func, struct svc_serv *serv,
458 struct svc_pool *pool)
460 struct svc_rqst *rqstp;
462 int have_oldmask = 0;
465 rqstp = kzalloc(sizeof(*rqstp), GFP_KERNEL);
469 init_waitqueue_head(&rqstp->rq_wait);
471 if (!(rqstp->rq_argp = kmalloc(serv->sv_xdrsize, GFP_KERNEL))
472 || !(rqstp->rq_resp = kmalloc(serv->sv_xdrsize, GFP_KERNEL))
473 || !svc_init_buffer(rqstp, serv->sv_bufsz))
476 serv->sv_nrthreads++;
477 spin_lock_bh(&pool->sp_lock);
478 pool->sp_nrthreads++;
479 list_add(&rqstp->rq_all, &pool->sp_all_threads);
480 spin_unlock_bh(&pool->sp_lock);
481 rqstp->rq_server = serv;
482 rqstp->rq_pool = pool;
484 if (serv->sv_nrpools > 1)
485 have_oldmask = svc_pool_map_set_cpumask(pool->sp_id, &oldmask);
487 error = kernel_thread((int (*)(void *)) func, rqstp, 0);
490 set_cpus_allowed(current, oldmask);
494 svc_sock_update_bufs(serv);
500 svc_exit_thread(rqstp);
505 * Create a thread in the default pool. Caller must hold BKL.
508 svc_create_thread(svc_thread_fn func, struct svc_serv *serv)
510 return __svc_create_thread(func, serv, &serv->sv_pools[0]);
514 * Choose a pool in which to create a new thread, for svc_set_num_threads
516 static inline struct svc_pool *
517 choose_pool(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
522 return &serv->sv_pools[(*state)++ % serv->sv_nrpools];
526 * Choose a thread to kill, for svc_set_num_threads
528 static inline struct task_struct *
529 choose_victim(struct svc_serv *serv, struct svc_pool *pool, unsigned int *state)
532 struct task_struct *task = NULL;
535 spin_lock_bh(&pool->sp_lock);
537 /* choose a pool in round-robin fashion */
538 for (i = 0; i < serv->sv_nrpools; i++) {
539 pool = &serv->sv_pools[--(*state) % serv->sv_nrpools];
540 spin_lock_bh(&pool->sp_lock);
541 if (!list_empty(&pool->sp_all_threads))
543 spin_unlock_bh(&pool->sp_lock);
549 if (!list_empty(&pool->sp_all_threads)) {
550 struct svc_rqst *rqstp;
553 * Remove from the pool->sp_all_threads list
554 * so we don't try to kill it again.
556 rqstp = list_entry(pool->sp_all_threads.next, struct svc_rqst, rq_all);
557 list_del_init(&rqstp->rq_all);
558 task = rqstp->rq_task;
560 spin_unlock_bh(&pool->sp_lock);
566 * Create or destroy enough new threads to make the number
567 * of threads the given number. If `pool' is non-NULL, applies
568 * only to threads in that pool, otherwise round-robins between
569 * all pools. Must be called with a svc_get() reference and
572 * Destroying threads relies on the service threads filling in
573 * rqstp->rq_task, which only the nfs ones do. Assumes the serv
574 * has been created using svc_create_pooled().
576 * Based on code that used to be in nfsd_svc() but tweaked
580 svc_set_num_threads(struct svc_serv *serv, struct svc_pool *pool, int nrservs)
582 struct task_struct *victim;
584 unsigned int state = serv->sv_nrthreads-1;
587 /* The -1 assumes caller has done a svc_get() */
588 nrservs -= (serv->sv_nrthreads-1);
590 spin_lock_bh(&pool->sp_lock);
591 nrservs -= pool->sp_nrthreads;
592 spin_unlock_bh(&pool->sp_lock);
595 /* create new threads */
596 while (nrservs > 0) {
598 __module_get(serv->sv_module);
599 error = __svc_create_thread(serv->sv_function, serv,
600 choose_pool(serv, pool, &state));
602 module_put(serv->sv_module);
606 /* destroy old threads */
607 while (nrservs < 0 &&
608 (victim = choose_victim(serv, pool, &state)) != NULL) {
609 send_sig(serv->sv_kill_signal, victim, 1);
617 * Called from a server thread as it's exiting. Caller must hold BKL.
620 svc_exit_thread(struct svc_rqst *rqstp)
622 struct svc_serv *serv = rqstp->rq_server;
623 struct svc_pool *pool = rqstp->rq_pool;
625 svc_release_buffer(rqstp);
626 kfree(rqstp->rq_resp);
627 kfree(rqstp->rq_argp);
628 kfree(rqstp->rq_auth_data);
630 spin_lock_bh(&pool->sp_lock);
631 pool->sp_nrthreads--;
632 list_del(&rqstp->rq_all);
633 spin_unlock_bh(&pool->sp_lock);
637 /* Release the server */
643 * Register an RPC service with the local portmapper.
644 * To unregister a service, call this routine with
645 * proto and port == 0.
648 svc_register(struct svc_serv *serv, int proto, unsigned short port)
650 struct svc_program *progp;
652 int i, error = 0, dummy;
654 progp = serv->sv_program;
656 dprintk("RPC: svc_register(%s, %s, %d)\n",
657 progp->pg_name, proto == IPPROTO_UDP? "udp" : "tcp", port);
660 clear_thread_flag(TIF_SIGPENDING);
662 for (i = 0; i < progp->pg_nvers; i++) {
663 if (progp->pg_vers[i] == NULL)
665 error = rpc_register(progp->pg_prog, i, proto, port, &dummy);
668 if (port && !dummy) {
675 spin_lock_irqsave(¤t->sighand->siglock, flags);
677 spin_unlock_irqrestore(¤t->sighand->siglock, flags);
684 * Process the RPC request.
687 svc_process(struct svc_rqst *rqstp)
689 struct svc_program *progp;
690 struct svc_version *versp = NULL; /* compiler food */
691 struct svc_procedure *procp = NULL;
692 struct kvec * argv = &rqstp->rq_arg.head[0];
693 struct kvec * resv = &rqstp->rq_res.head[0];
694 struct svc_serv *serv = rqstp->rq_server;
697 u32 dir, prog, vers, proc;
698 __be32 auth_stat, rpc_stat;
700 __be32 *accept_statp;
702 rpc_stat = rpc_success;
704 if (argv->iov_len < 6*4)
707 /* setup response xdr_buf.
708 * Initially it has just one page
710 svc_take_page(rqstp); /* must succeed */
711 resv->iov_base = page_address(rqstp->rq_respages[0]);
713 rqstp->rq_res.pages = rqstp->rq_respages+1;
714 rqstp->rq_res.len = 0;
715 rqstp->rq_res.page_base = 0;
716 rqstp->rq_res.page_len = 0;
717 rqstp->rq_res.buflen = PAGE_SIZE;
718 rqstp->rq_res.tail[0].iov_base = NULL;
719 rqstp->rq_res.tail[0].iov_len = 0;
720 /* Will be turned off only in gss privacy case: */
721 rqstp->rq_sendfile_ok = 1;
722 /* tcp needs a space for the record length... */
723 if (rqstp->rq_prot == IPPROTO_TCP)
726 rqstp->rq_xid = svc_getu32(argv);
727 svc_putu32(resv, rqstp->rq_xid);
729 dir = svc_getnl(argv);
730 vers = svc_getnl(argv);
732 /* First words of reply: */
733 svc_putnl(resv, 1); /* REPLY */
735 if (dir != 0) /* direction != CALL */
737 if (vers != 2) /* RPC version number */
740 /* Save position in case we later decide to reject: */
741 accept_statp = resv->iov_base + resv->iov_len;
743 svc_putnl(resv, 0); /* ACCEPT */
745 rqstp->rq_prog = prog = svc_getnl(argv); /* program number */
746 rqstp->rq_vers = vers = svc_getnl(argv); /* version number */
747 rqstp->rq_proc = proc = svc_getnl(argv); /* procedure number */
749 progp = serv->sv_program;
751 for (progp = serv->sv_program; progp; progp = progp->pg_next)
752 if (prog == progp->pg_prog)
756 * Decode auth data, and add verifier to reply buffer.
757 * We do this before anything else in order to get a decent
760 auth_res = svc_authenticate(rqstp, &auth_stat);
761 /* Also give the program a chance to reject this call: */
762 if (auth_res == SVC_OK && progp) {
763 auth_stat = rpc_autherr_badcred;
764 auth_res = progp->pg_authenticate(rqstp);
770 rpc_stat = rpc_garbage_args;
773 rpc_stat = rpc_system_err;
786 if (vers >= progp->pg_nvers ||
787 !(versp = progp->pg_vers[vers]))
790 procp = versp->vs_proc + proc;
791 if (proc >= versp->vs_nproc || !procp->pc_func)
793 rqstp->rq_server = serv;
794 rqstp->rq_procinfo = procp;
796 /* Syntactic check complete */
797 serv->sv_stats->rpccnt++;
799 /* Build the reply header. */
800 statp = resv->iov_base +resv->iov_len;
801 svc_putnl(resv, RPC_SUCCESS);
803 /* Bump per-procedure stats counter */
806 /* Initialize storage for argp and resp */
807 memset(rqstp->rq_argp, 0, procp->pc_argsize);
808 memset(rqstp->rq_resp, 0, procp->pc_ressize);
810 /* un-reserve some of the out-queue now that we have a
811 * better idea of reply size
813 if (procp->pc_xdrressize)
814 svc_reserve(rqstp, procp->pc_xdrressize<<2);
816 /* Call the function that processes the request. */
817 if (!versp->vs_dispatch) {
818 /* Decode arguments */
819 xdr = procp->pc_decode;
820 if (xdr && !xdr(rqstp, argv->iov_base, rqstp->rq_argp))
823 *statp = procp->pc_func(rqstp, rqstp->rq_argp, rqstp->rq_resp);
826 if (*statp == rpc_success && (xdr = procp->pc_encode)
827 && !xdr(rqstp, resv->iov_base+resv->iov_len, rqstp->rq_resp)) {
828 dprintk("svc: failed to encode reply\n");
829 /* serv->sv_stats->rpcsystemerr++; */
830 *statp = rpc_system_err;
833 dprintk("svc: calling dispatcher\n");
834 if (!versp->vs_dispatch(rqstp, statp)) {
835 /* Release reply info */
836 if (procp->pc_release)
837 procp->pc_release(rqstp, NULL, rqstp->rq_resp);
842 /* Check RPC status result */
843 if (*statp != rpc_success)
844 resv->iov_len = ((void*)statp) - resv->iov_base + 4;
846 /* Release reply info */
847 if (procp->pc_release)
848 procp->pc_release(rqstp, NULL, rqstp->rq_resp);
850 if (procp->pc_encode == NULL)
854 if (svc_authorise(rqstp))
856 return svc_send(rqstp);
859 svc_authorise(rqstp); /* doesn't hurt to call this twice */
860 dprintk("svc: svc_process dropit\n");
866 printk("svc: short len %Zd, dropping request\n", argv->iov_len);
868 goto dropit; /* drop request */
872 printk("svc: bad direction %d, dropping request\n", dir);
874 serv->sv_stats->rpcbadfmt++;
875 goto dropit; /* drop request */
878 serv->sv_stats->rpcbadfmt++;
879 svc_putnl(resv, 1); /* REJECT */
880 svc_putnl(resv, 0); /* RPC_MISMATCH */
881 svc_putnl(resv, 2); /* Only RPCv2 supported */
886 dprintk("svc: authentication failed (%d)\n", ntohl(auth_stat));
887 serv->sv_stats->rpcbadauth++;
888 /* Restore write pointer to location of accept status: */
889 xdr_ressize_check(rqstp, accept_statp);
890 svc_putnl(resv, 1); /* REJECT */
891 svc_putnl(resv, 1); /* AUTH_ERROR */
892 svc_putnl(resv, ntohl(auth_stat)); /* status */
896 dprintk("svc: unknown program %d\n", prog);
897 serv->sv_stats->rpcbadfmt++;
898 svc_putnl(resv, RPC_PROG_UNAVAIL);
903 printk("svc: unknown version (%d)\n", vers);
905 serv->sv_stats->rpcbadfmt++;
906 svc_putnl(resv, RPC_PROG_MISMATCH);
907 svc_putnl(resv, progp->pg_lovers);
908 svc_putnl(resv, progp->pg_hivers);
913 printk("svc: unknown procedure (%d)\n", proc);
915 serv->sv_stats->rpcbadfmt++;
916 svc_putnl(resv, RPC_PROC_UNAVAIL);
921 printk("svc: failed to decode args\n");
923 rpc_stat = rpc_garbage_args;
925 serv->sv_stats->rpcbadfmt++;
926 svc_putnl(resv, ntohl(rpc_stat));