LockD: manage garbage collection timeout per networks namespace
[linux-2.6-block.git] / fs / lockd / host.c
1 /*
2  * linux/fs/lockd/host.c
3  *
4  * Management for NLM peer hosts. The nlm_host struct is shared
5  * between client and server implementation. The only reason to
6  * do so is to reduce code bloat.
7  *
8  * Copyright (C) 1996, Olaf Kirch <okir@monad.swb.de>
9  */
10
11 #include <linux/types.h>
12 #include <linux/slab.h>
13 #include <linux/in.h>
14 #include <linux/in6.h>
15 #include <linux/sunrpc/clnt.h>
16 #include <linux/sunrpc/svc.h>
17 #include <linux/lockd/lockd.h>
18 #include <linux/mutex.h>
19
20 #include <linux/sunrpc/svc_xprt.h>
21
22 #include <net/ipv6.h>
23
24 #include "netns.h"
25
26 #define NLMDBG_FACILITY         NLMDBG_HOSTCACHE
27 #define NLM_HOST_NRHASH         32
28 #define NLM_HOST_REBIND         (60 * HZ)
29 #define NLM_HOST_EXPIRE         (300 * HZ)
30 #define NLM_HOST_COLLECT        (120 * HZ)
31
32 static struct hlist_head        nlm_server_hosts[NLM_HOST_NRHASH];
33 static struct hlist_head        nlm_client_hosts[NLM_HOST_NRHASH];
34
35 #define for_each_host(host, pos, chain, table) \
36         for ((chain) = (table); \
37              (chain) < (table) + NLM_HOST_NRHASH; ++(chain)) \
38                 hlist_for_each_entry((host), (pos), (chain), h_hash)
39
40 #define for_each_host_safe(host, pos, next, chain, table) \
41         for ((chain) = (table); \
42              (chain) < (table) + NLM_HOST_NRHASH; ++(chain)) \
43                 hlist_for_each_entry_safe((host), (pos), (next), \
44                                                 (chain), h_hash)
45
46 static unsigned long            nrhosts;
47 static DEFINE_MUTEX(nlm_host_mutex);
48
49 static void                     nlm_gc_hosts(struct net *net);
50
51 struct nlm_lookup_host_info {
52         const int               server;         /* search for server|client */
53         const struct sockaddr   *sap;           /* address to search for */
54         const size_t            salen;          /* it's length */
55         const unsigned short    protocol;       /* transport to search for*/
56         const u32               version;        /* NLM version to search for */
57         const char              *hostname;      /* remote's hostname */
58         const size_t            hostname_len;   /* it's length */
59         const int               noresvport;     /* use non-priv port */
60         struct net              *net;           /* network namespace to bind */
61 };
62
63 /*
64  * Hash function must work well on big- and little-endian platforms
65  */
66 static unsigned int __nlm_hash32(const __be32 n)
67 {
68         unsigned int hash = (__force u32)n ^ ((__force u32)n >> 16);
69         return hash ^ (hash >> 8);
70 }
71
72 static unsigned int __nlm_hash_addr4(const struct sockaddr *sap)
73 {
74         const struct sockaddr_in *sin = (struct sockaddr_in *)sap;
75         return __nlm_hash32(sin->sin_addr.s_addr);
76 }
77
78 static unsigned int __nlm_hash_addr6(const struct sockaddr *sap)
79 {
80         const struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)sap;
81         const struct in6_addr addr = sin6->sin6_addr;
82         return __nlm_hash32(addr.s6_addr32[0]) ^
83                __nlm_hash32(addr.s6_addr32[1]) ^
84                __nlm_hash32(addr.s6_addr32[2]) ^
85                __nlm_hash32(addr.s6_addr32[3]);
86 }
87
88 static unsigned int nlm_hash_address(const struct sockaddr *sap)
89 {
90         unsigned int hash;
91
92         switch (sap->sa_family) {
93         case AF_INET:
94                 hash = __nlm_hash_addr4(sap);
95                 break;
96         case AF_INET6:
97                 hash = __nlm_hash_addr6(sap);
98                 break;
99         default:
100                 hash = 0;
101         }
102         return hash & (NLM_HOST_NRHASH - 1);
103 }
104
105 /*
106  * Allocate and initialize an nlm_host.  Common to both client and server.
107  */
108 static struct nlm_host *nlm_alloc_host(struct nlm_lookup_host_info *ni,
109                                        struct nsm_handle *nsm)
110 {
111         struct nlm_host *host = NULL;
112         unsigned long now = jiffies;
113
114         if (nsm != NULL)
115                 atomic_inc(&nsm->sm_count);
116         else {
117                 host = NULL;
118                 nsm = nsm_get_handle(ni->sap, ni->salen,
119                                         ni->hostname, ni->hostname_len);
120                 if (unlikely(nsm == NULL)) {
121                         dprintk("lockd: %s failed; no nsm handle\n",
122                                 __func__);
123                         goto out;
124                 }
125         }
126
127         host = kmalloc(sizeof(*host), GFP_KERNEL);
128         if (unlikely(host == NULL)) {
129                 dprintk("lockd: %s failed; no memory\n", __func__);
130                 nsm_release(nsm);
131                 goto out;
132         }
133
134         memcpy(nlm_addr(host), ni->sap, ni->salen);
135         host->h_addrlen    = ni->salen;
136         rpc_set_port(nlm_addr(host), 0);
137         host->h_srcaddrlen = 0;
138
139         host->h_rpcclnt    = NULL;
140         host->h_name       = nsm->sm_name;
141         host->h_version    = ni->version;
142         host->h_proto      = ni->protocol;
143         host->h_reclaiming = 0;
144         host->h_server     = ni->server;
145         host->h_noresvport = ni->noresvport;
146         host->h_inuse      = 0;
147         init_waitqueue_head(&host->h_gracewait);
148         init_rwsem(&host->h_rwsem);
149         host->h_state      = 0;
150         host->h_nsmstate   = 0;
151         host->h_pidcount   = 0;
152         atomic_set(&host->h_count, 1);
153         mutex_init(&host->h_mutex);
154         host->h_nextrebind = now + NLM_HOST_REBIND;
155         host->h_expires    = now + NLM_HOST_EXPIRE;
156         INIT_LIST_HEAD(&host->h_lockowners);
157         spin_lock_init(&host->h_lock);
158         INIT_LIST_HEAD(&host->h_granted);
159         INIT_LIST_HEAD(&host->h_reclaim);
160         host->h_nsmhandle  = nsm;
161         host->h_addrbuf    = nsm->sm_addrbuf;
162         host->net          = ni->net;
163
164 out:
165         return host;
166 }
167
168 /*
169  * Destroy an nlm_host and free associated resources
170  *
171  * Caller must hold nlm_host_mutex.
172  */
173 static void nlm_destroy_host_locked(struct nlm_host *host)
174 {
175         struct rpc_clnt *clnt;
176
177         dprintk("lockd: destroy host %s\n", host->h_name);
178
179         BUG_ON(!list_empty(&host->h_lockowners));
180         BUG_ON(atomic_read(&host->h_count));
181
182         hlist_del_init(&host->h_hash);
183
184         nsm_unmonitor(host);
185         nsm_release(host->h_nsmhandle);
186
187         clnt = host->h_rpcclnt;
188         if (clnt != NULL)
189                 rpc_shutdown_client(clnt);
190         kfree(host);
191
192         nrhosts--;
193 }
194
195 /**
196  * nlmclnt_lookup_host - Find an NLM host handle matching a remote server
197  * @sap: network address of server
198  * @salen: length of server address
199  * @protocol: transport protocol to use
200  * @version: NLM protocol version
201  * @hostname: '\0'-terminated hostname of server
202  * @noresvport: 1 if non-privileged port should be used
203  *
204  * Returns an nlm_host structure that matches the passed-in
205  * [server address, transport protocol, NLM version, server hostname].
206  * If one doesn't already exist in the host cache, a new handle is
207  * created and returned.
208  */
209 struct nlm_host *nlmclnt_lookup_host(const struct sockaddr *sap,
210                                      const size_t salen,
211                                      const unsigned short protocol,
212                                      const u32 version,
213                                      const char *hostname,
214                                      int noresvport,
215                                      struct net *net)
216 {
217         struct nlm_lookup_host_info ni = {
218                 .server         = 0,
219                 .sap            = sap,
220                 .salen          = salen,
221                 .protocol       = protocol,
222                 .version        = version,
223                 .hostname       = hostname,
224                 .hostname_len   = strlen(hostname),
225                 .noresvport     = noresvport,
226                 .net            = net,
227         };
228         struct hlist_head *chain;
229         struct hlist_node *pos;
230         struct nlm_host *host;
231         struct nsm_handle *nsm = NULL;
232
233         dprintk("lockd: %s(host='%s', vers=%u, proto=%s)\n", __func__,
234                         (hostname ? hostname : "<none>"), version,
235                         (protocol == IPPROTO_UDP ? "udp" : "tcp"));
236
237         mutex_lock(&nlm_host_mutex);
238
239         chain = &nlm_client_hosts[nlm_hash_address(sap)];
240         hlist_for_each_entry(host, pos, chain, h_hash) {
241                 if (host->net != net)
242                         continue;
243                 if (!rpc_cmp_addr(nlm_addr(host), sap))
244                         continue;
245
246                 /* Same address. Share an NSM handle if we already have one */
247                 if (nsm == NULL)
248                         nsm = host->h_nsmhandle;
249
250                 if (host->h_proto != protocol)
251                         continue;
252                 if (host->h_version != version)
253                         continue;
254
255                 nlm_get_host(host);
256                 dprintk("lockd: %s found host %s (%s)\n", __func__,
257                         host->h_name, host->h_addrbuf);
258                 goto out;
259         }
260
261         host = nlm_alloc_host(&ni, nsm);
262         if (unlikely(host == NULL))
263                 goto out;
264
265         hlist_add_head(&host->h_hash, chain);
266         nrhosts++;
267
268         dprintk("lockd: %s created host %s (%s)\n", __func__,
269                 host->h_name, host->h_addrbuf);
270
271 out:
272         mutex_unlock(&nlm_host_mutex);
273         return host;
274 }
275
276 /**
277  * nlmclnt_release_host - release client nlm_host
278  * @host: nlm_host to release
279  *
280  */
281 void nlmclnt_release_host(struct nlm_host *host)
282 {
283         if (host == NULL)
284                 return;
285
286         dprintk("lockd: release client host %s\n", host->h_name);
287
288         BUG_ON(atomic_read(&host->h_count) < 0);
289         BUG_ON(host->h_server);
290
291         if (atomic_dec_and_test(&host->h_count)) {
292                 BUG_ON(!list_empty(&host->h_lockowners));
293                 BUG_ON(!list_empty(&host->h_granted));
294                 BUG_ON(!list_empty(&host->h_reclaim));
295
296                 mutex_lock(&nlm_host_mutex);
297                 nlm_destroy_host_locked(host);
298                 mutex_unlock(&nlm_host_mutex);
299         }
300 }
301
302 /**
303  * nlmsvc_lookup_host - Find an NLM host handle matching a remote client
304  * @rqstp: incoming NLM request
305  * @hostname: name of client host
306  * @hostname_len: length of client hostname
307  *
308  * Returns an nlm_host structure that matches the [client address,
309  * transport protocol, NLM version, client hostname] of the passed-in
310  * NLM request.  If one doesn't already exist in the host cache, a
311  * new handle is created and returned.
312  *
313  * Before possibly creating a new nlm_host, construct a sockaddr
314  * for a specific source address in case the local system has
315  * multiple network addresses.  The family of the address in
316  * rq_daddr is guaranteed to be the same as the family of the
317  * address in rq_addr, so it's safe to use the same family for
318  * the source address.
319  */
320 struct nlm_host *nlmsvc_lookup_host(const struct svc_rqst *rqstp,
321                                     const char *hostname,
322                                     const size_t hostname_len)
323 {
324         struct hlist_head *chain;
325         struct hlist_node *pos;
326         struct nlm_host *host = NULL;
327         struct nsm_handle *nsm = NULL;
328         struct sockaddr *src_sap = svc_daddr(rqstp);
329         size_t src_len = rqstp->rq_daddrlen;
330         struct net *net = rqstp->rq_xprt->xpt_net;
331         struct nlm_lookup_host_info ni = {
332                 .server         = 1,
333                 .sap            = svc_addr(rqstp),
334                 .salen          = rqstp->rq_addrlen,
335                 .protocol       = rqstp->rq_prot,
336                 .version        = rqstp->rq_vers,
337                 .hostname       = hostname,
338                 .hostname_len   = hostname_len,
339                 .net            = net,
340         };
341         struct lockd_net *ln = net_generic(net, lockd_net_id);
342
343         dprintk("lockd: %s(host='%*s', vers=%u, proto=%s)\n", __func__,
344                         (int)hostname_len, hostname, rqstp->rq_vers,
345                         (rqstp->rq_prot == IPPROTO_UDP ? "udp" : "tcp"));
346
347         mutex_lock(&nlm_host_mutex);
348
349         if (time_after_eq(jiffies, ln->next_gc))
350                 nlm_gc_hosts(net);
351
352         chain = &nlm_server_hosts[nlm_hash_address(ni.sap)];
353         hlist_for_each_entry(host, pos, chain, h_hash) {
354                 if (host->net != net)
355                         continue;
356                 if (!rpc_cmp_addr(nlm_addr(host), ni.sap))
357                         continue;
358
359                 /* Same address. Share an NSM handle if we already have one */
360                 if (nsm == NULL)
361                         nsm = host->h_nsmhandle;
362
363                 if (host->h_proto != ni.protocol)
364                         continue;
365                 if (host->h_version != ni.version)
366                         continue;
367                 if (!rpc_cmp_addr(nlm_srcaddr(host), src_sap))
368                         continue;
369
370                 /* Move to head of hash chain. */
371                 hlist_del(&host->h_hash);
372                 hlist_add_head(&host->h_hash, chain);
373
374                 nlm_get_host(host);
375                 dprintk("lockd: %s found host %s (%s)\n",
376                         __func__, host->h_name, host->h_addrbuf);
377                 goto out;
378         }
379
380         host = nlm_alloc_host(&ni, nsm);
381         if (unlikely(host == NULL))
382                 goto out;
383
384         memcpy(nlm_srcaddr(host), src_sap, src_len);
385         host->h_srcaddrlen = src_len;
386         hlist_add_head(&host->h_hash, chain);
387         nrhosts++;
388
389         dprintk("lockd: %s created host %s (%s)\n",
390                 __func__, host->h_name, host->h_addrbuf);
391
392 out:
393         mutex_unlock(&nlm_host_mutex);
394         return host;
395 }
396
397 /**
398  * nlmsvc_release_host - release server nlm_host
399  * @host: nlm_host to release
400  *
401  * Host is destroyed later in nlm_gc_host().
402  */
403 void nlmsvc_release_host(struct nlm_host *host)
404 {
405         if (host == NULL)
406                 return;
407
408         dprintk("lockd: release server host %s\n", host->h_name);
409
410         BUG_ON(atomic_read(&host->h_count) < 0);
411         BUG_ON(!host->h_server);
412         atomic_dec(&host->h_count);
413 }
414
415 /*
416  * Create the NLM RPC client for an NLM peer
417  */
418 struct rpc_clnt *
419 nlm_bind_host(struct nlm_host *host)
420 {
421         struct rpc_clnt *clnt;
422
423         dprintk("lockd: nlm_bind_host %s (%s)\n",
424                         host->h_name, host->h_addrbuf);
425
426         /* Lock host handle */
427         mutex_lock(&host->h_mutex);
428
429         /* If we've already created an RPC client, check whether
430          * RPC rebind is required
431          */
432         if ((clnt = host->h_rpcclnt) != NULL) {
433                 if (time_after_eq(jiffies, host->h_nextrebind)) {
434                         rpc_force_rebind(clnt);
435                         host->h_nextrebind = jiffies + NLM_HOST_REBIND;
436                         dprintk("lockd: next rebind in %lu jiffies\n",
437                                         host->h_nextrebind - jiffies);
438                 }
439         } else {
440                 unsigned long increment = nlmsvc_timeout;
441                 struct rpc_timeout timeparms = {
442                         .to_initval     = increment,
443                         .to_increment   = increment,
444                         .to_maxval      = increment * 6UL,
445                         .to_retries     = 5U,
446                 };
447                 struct rpc_create_args args = {
448                         .net            = host->net,
449                         .protocol       = host->h_proto,
450                         .address        = nlm_addr(host),
451                         .addrsize       = host->h_addrlen,
452                         .timeout        = &timeparms,
453                         .servername     = host->h_name,
454                         .program        = &nlm_program,
455                         .version        = host->h_version,
456                         .authflavor     = RPC_AUTH_UNIX,
457                         .flags          = (RPC_CLNT_CREATE_NOPING |
458                                            RPC_CLNT_CREATE_AUTOBIND),
459                 };
460
461                 /*
462                  * lockd retries server side blocks automatically so we want
463                  * those to be soft RPC calls. Client side calls need to be
464                  * hard RPC tasks.
465                  */
466                 if (!host->h_server)
467                         args.flags |= RPC_CLNT_CREATE_HARDRTRY;
468                 if (host->h_noresvport)
469                         args.flags |= RPC_CLNT_CREATE_NONPRIVPORT;
470                 if (host->h_srcaddrlen)
471                         args.saddress = nlm_srcaddr(host);
472
473                 clnt = rpc_create(&args);
474                 if (!IS_ERR(clnt))
475                         host->h_rpcclnt = clnt;
476                 else {
477                         printk("lockd: couldn't create RPC handle for %s\n", host->h_name);
478                         clnt = NULL;
479                 }
480         }
481
482         mutex_unlock(&host->h_mutex);
483         return clnt;
484 }
485
486 /*
487  * Force a portmap lookup of the remote lockd port
488  */
489 void
490 nlm_rebind_host(struct nlm_host *host)
491 {
492         dprintk("lockd: rebind host %s\n", host->h_name);
493         if (host->h_rpcclnt && time_after_eq(jiffies, host->h_nextrebind)) {
494                 rpc_force_rebind(host->h_rpcclnt);
495                 host->h_nextrebind = jiffies + NLM_HOST_REBIND;
496         }
497 }
498
499 /*
500  * Increment NLM host count
501  */
502 struct nlm_host * nlm_get_host(struct nlm_host *host)
503 {
504         if (host) {
505                 dprintk("lockd: get host %s\n", host->h_name);
506                 atomic_inc(&host->h_count);
507                 host->h_expires = jiffies + NLM_HOST_EXPIRE;
508         }
509         return host;
510 }
511
512 static struct nlm_host *next_host_state(struct hlist_head *cache,
513                                         struct nsm_handle *nsm,
514                                         const struct nlm_reboot *info)
515 {
516         struct nlm_host *host;
517         struct hlist_head *chain;
518         struct hlist_node *pos;
519
520         mutex_lock(&nlm_host_mutex);
521         for_each_host(host, pos, chain, cache) {
522                 if (host->h_nsmhandle == nsm
523                     && host->h_nsmstate != info->state) {
524                         host->h_nsmstate = info->state;
525                         host->h_state++;
526
527                         nlm_get_host(host);
528                         mutex_unlock(&nlm_host_mutex);
529                         return host;
530                 }
531         }
532
533         mutex_unlock(&nlm_host_mutex);
534         return NULL;
535 }
536
537 /**
538  * nlm_host_rebooted - Release all resources held by rebooted host
539  * @info: pointer to decoded results of NLM_SM_NOTIFY call
540  *
541  * We were notified that the specified host has rebooted.  Release
542  * all resources held by that peer.
543  */
544 void nlm_host_rebooted(const struct nlm_reboot *info)
545 {
546         struct nsm_handle *nsm;
547         struct nlm_host *host;
548
549         nsm = nsm_reboot_lookup(info);
550         if (unlikely(nsm == NULL))
551                 return;
552
553         /* Mark all hosts tied to this NSM state as having rebooted.
554          * We run the loop repeatedly, because we drop the host table
555          * lock for this.
556          * To avoid processing a host several times, we match the nsmstate.
557          */
558         while ((host = next_host_state(nlm_server_hosts, nsm, info)) != NULL) {
559                 nlmsvc_free_host_resources(host);
560                 nlmsvc_release_host(host);
561         }
562         while ((host = next_host_state(nlm_client_hosts, nsm, info)) != NULL) {
563                 nlmclnt_recovery(host);
564                 nlmclnt_release_host(host);
565         }
566
567         nsm_release(nsm);
568 }
569
570 void
571 nlm_shutdown_hosts_net(struct net *net)
572 {
573         struct hlist_head *chain;
574         struct hlist_node *pos;
575         struct nlm_host *host;
576
577         dprintk("lockd: shutting down host module\n");
578         mutex_lock(&nlm_host_mutex);
579
580         /* First, make all hosts eligible for gc */
581         dprintk("lockd: nuking all hosts...\n");
582         for_each_host(host, pos, chain, nlm_server_hosts) {
583                 if (net && host->net != net)
584                         continue;
585                 host->h_expires = jiffies - 1;
586                 if (host->h_rpcclnt) {
587                         rpc_shutdown_client(host->h_rpcclnt);
588                         host->h_rpcclnt = NULL;
589                 }
590         }
591
592         /* Then, perform a garbage collection pass */
593         nlm_gc_hosts(net);
594         mutex_unlock(&nlm_host_mutex);
595 }
596
597 /*
598  * Shut down the hosts module.
599  * Note that this routine is called only at server shutdown time.
600  */
601 void
602 nlm_shutdown_hosts(void)
603 {
604         struct hlist_head *chain;
605         struct hlist_node *pos;
606         struct nlm_host *host;
607
608         nlm_shutdown_hosts_net(NULL);
609
610         /* complain if any hosts are left */
611         if (nrhosts != 0) {
612                 printk(KERN_WARNING "lockd: couldn't shutdown host module!\n");
613                 dprintk("lockd: %lu hosts left:\n", nrhosts);
614                 for_each_host(host, pos, chain, nlm_server_hosts) {
615                         dprintk("       %s (cnt %d use %d exp %ld net %p)\n",
616                                 host->h_name, atomic_read(&host->h_count),
617                                 host->h_inuse, host->h_expires, host->net);
618                 }
619         }
620 }
621
622 /*
623  * Garbage collect any unused NLM hosts.
624  * This GC combines reference counting for async operations with
625  * mark & sweep for resources held by remote clients.
626  */
627 static void
628 nlm_gc_hosts(struct net *net)
629 {
630         struct hlist_head *chain;
631         struct hlist_node *pos, *next;
632         struct nlm_host *host;
633
634         dprintk("lockd: host garbage collection for net %p\n", net);
635         for_each_host(host, pos, chain, nlm_server_hosts) {
636                 if (net && host->net != net)
637                         continue;
638                 host->h_inuse = 0;
639         }
640
641         /* Mark all hosts that hold locks, blocks or shares */
642         nlmsvc_mark_resources(net);
643
644         for_each_host_safe(host, pos, next, chain, nlm_server_hosts) {
645                 if (net && host->net != net)
646                         continue;
647                 if (atomic_read(&host->h_count) || host->h_inuse
648                  || time_before(jiffies, host->h_expires)) {
649                         dprintk("nlm_gc_hosts skipping %s "
650                                 "(cnt %d use %d exp %ld net %p)\n",
651                                 host->h_name, atomic_read(&host->h_count),
652                                 host->h_inuse, host->h_expires, host->net);
653                         continue;
654                 }
655                 nlm_destroy_host_locked(host);
656         }
657
658         if (net) {
659                 struct lockd_net *ln = net_generic(net, lockd_net_id);
660
661                 ln->next_gc = jiffies + NLM_HOST_COLLECT;
662         }
663 }