blob: 2cbac34a55da1ac6cebeefabe2bdfdd5b883b9f7 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002 * Request reply cache. This is currently a global cache, but this may
3 * change in the future and be a per-client cache.
4 *
5 * This code is heavily inspired by the 44BSD implementation, although
6 * it does things a bit differently.
7 *
8 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
9 */
10
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090011#include <linux/slab.h>
12
Boaz Harrosh9a74af22009-12-03 20:30:56 +020013#include "nfsd.h"
14#include "cache.h"
Linus Torvalds1da177e2005-04-16 15:20:36 -070015
16/* Size of reply cache. Common values are:
17 * 4.3BSD: 128
18 * 4.4BSD: 256
19 * Solaris2: 1024
20 * DEC Unix: 512-4096
21 */
22#define CACHESIZE 1024
23#define HASHSIZE 64
Linus Torvalds1da177e2005-04-16 15:20:36 -070024
Greg Banksfca42172009-04-01 07:28:13 +110025static struct hlist_head * cache_hash;
Linus Torvalds1da177e2005-04-16 15:20:36 -070026static struct list_head lru_head;
27static int cache_disabled = 1;
28
Greg Banksfca42172009-04-01 07:28:13 +110029/*
30 * Calculate the hash index from an XID.
31 */
32static inline u32 request_hash(u32 xid)
33{
34 u32 h = xid;
35 h ^= (xid >> 24);
36 return h & (HASHSIZE-1);
37}
38
Linus Torvalds1da177e2005-04-16 15:20:36 -070039static int nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
40
Greg Banksfca42172009-04-01 07:28:13 +110041/*
Linus Torvalds1da177e2005-04-16 15:20:36 -070042 * locking for the reply cache:
43 * A cache entry is "single use" if c_state == RC_INPROG
44 * Otherwise, it when accessing _prev or _next, the lock must be held.
45 */
46static DEFINE_SPINLOCK(cache_lock);
47
J. Bruce Fieldsd5c34282007-11-09 14:10:56 -050048int nfsd_reply_cache_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -070049{
50 struct svc_cacherep *rp;
51 int i;
52
53 INIT_LIST_HEAD(&lru_head);
54 i = CACHESIZE;
J. Bruce Fieldsd5c34282007-11-09 14:10:56 -050055 while (i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -070056 rp = kmalloc(sizeof(*rp), GFP_KERNEL);
J. Bruce Fieldsd5c34282007-11-09 14:10:56 -050057 if (!rp)
58 goto out_nomem;
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 list_add(&rp->c_lru, &lru_head);
60 rp->c_state = RC_UNUSED;
61 rp->c_type = RC_NOCACHE;
62 INIT_HLIST_NODE(&rp->c_hash);
63 i--;
64 }
65
Greg Banksfca42172009-04-01 07:28:13 +110066 cache_hash = kcalloc (HASHSIZE, sizeof(struct hlist_head), GFP_KERNEL);
67 if (!cache_hash)
J. Bruce Fieldsd5c34282007-11-09 14:10:56 -050068 goto out_nomem;
Linus Torvalds1da177e2005-04-16 15:20:36 -070069
70 cache_disabled = 0;
J. Bruce Fieldsd5c34282007-11-09 14:10:56 -050071 return 0;
72out_nomem:
73 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
74 nfsd_reply_cache_shutdown();
75 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -070076}
77
J. Bruce Fieldsd5c34282007-11-09 14:10:56 -050078void nfsd_reply_cache_shutdown(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -070079{
80 struct svc_cacherep *rp;
81
82 while (!list_empty(&lru_head)) {
83 rp = list_entry(lru_head.next, struct svc_cacherep, c_lru);
84 if (rp->c_state == RC_DONE && rp->c_type == RC_REPLBUFF)
85 kfree(rp->c_replvec.iov_base);
86 list_del(&rp->c_lru);
87 kfree(rp);
88 }
89
90 cache_disabled = 1;
91
Greg Banksfca42172009-04-01 07:28:13 +110092 kfree (cache_hash);
93 cache_hash = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -070094}
95
96/*
97 * Move cache entry to end of LRU list
98 */
99static void
100lru_put_end(struct svc_cacherep *rp)
101{
Akinobu Mitaf1166292006-06-26 00:24:46 -0700102 list_move_tail(&rp->c_lru, &lru_head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103}
104
105/*
106 * Move a cache entry from one hash list to another
107 */
108static void
109hash_refile(struct svc_cacherep *rp)
110{
111 hlist_del_init(&rp->c_hash);
Greg Banksfca42172009-04-01 07:28:13 +1100112 hlist_add_head(&rp->c_hash, cache_hash + request_hash(rp->c_xid));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113}
114
115/*
116 * Try to find an entry matching the current call in the cache. When none
117 * is found, we grab the oldest unlocked entry off the LRU list.
118 * Note that no operation within the loop may sleep.
119 */
120int
J. Bruce Fields10910062011-01-24 12:11:02 -0500121nfsd_cache_lookup(struct svc_rqst *rqstp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122{
123 struct hlist_node *hn;
124 struct hlist_head *rh;
125 struct svc_cacherep *rp;
Al Viroc7afef12006-10-19 23:29:02 -0700126 __be32 xid = rqstp->rq_xid;
127 u32 proto = rqstp->rq_prot,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128 vers = rqstp->rq_vers,
129 proc = rqstp->rq_proc;
130 unsigned long age;
J. Bruce Fields10910062011-01-24 12:11:02 -0500131 int type = rqstp->rq_cachetype;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132 int rtn;
133
134 rqstp->rq_cacherep = NULL;
135 if (cache_disabled || type == RC_NOCACHE) {
136 nfsdstats.rcnocache++;
137 return RC_DOIT;
138 }
139
140 spin_lock(&cache_lock);
141 rtn = RC_DOIT;
142
Greg Banksfca42172009-04-01 07:28:13 +1100143 rh = &cache_hash[request_hash(xid)];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144 hlist_for_each_entry(rp, hn, rh, c_hash) {
145 if (rp->c_state != RC_UNUSED &&
146 xid == rp->c_xid && proc == rp->c_proc &&
147 proto == rp->c_prot && vers == rp->c_vers &&
148 time_before(jiffies, rp->c_timestamp + 120*HZ) &&
149 memcmp((char*)&rqstp->rq_addr, (char*)&rp->c_addr, sizeof(rp->c_addr))==0) {
150 nfsdstats.rchits++;
151 goto found_entry;
152 }
153 }
154 nfsdstats.rcmisses++;
155
156 /* This loop shouldn't take more than a few iterations normally */
157 {
158 int safe = 0;
159 list_for_each_entry(rp, &lru_head, c_lru) {
160 if (rp->c_state != RC_INPROG)
161 break;
162 if (safe++ > CACHESIZE) {
163 printk("nfsd: loop in repcache LRU list\n");
164 cache_disabled = 1;
165 goto out;
166 }
167 }
168 }
169
Greg Bankscf0a5862009-04-01 07:28:15 +1100170 /* All entries on the LRU are in-progress. This should not happen */
171 if (&rp->c_lru == &lru_head) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172 static int complaints;
173
174 printk(KERN_WARNING "nfsd: all repcache entries locked!\n");
175 if (++complaints > 5) {
176 printk(KERN_WARNING "nfsd: disabling repcache.\n");
177 cache_disabled = 1;
178 }
179 goto out;
180 }
181
182 rqstp->rq_cacherep = rp;
183 rp->c_state = RC_INPROG;
184 rp->c_xid = xid;
185 rp->c_proc = proc;
Chuck Lever27459f02007-02-12 00:53:34 -0800186 memcpy(&rp->c_addr, svc_addr_in(rqstp), sizeof(rp->c_addr));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187 rp->c_prot = proto;
188 rp->c_vers = vers;
189 rp->c_timestamp = jiffies;
190
191 hash_refile(rp);
192
193 /* release any buffer */
194 if (rp->c_type == RC_REPLBUFF) {
195 kfree(rp->c_replvec.iov_base);
196 rp->c_replvec.iov_base = NULL;
197 }
198 rp->c_type = RC_NOCACHE;
199 out:
200 spin_unlock(&cache_lock);
201 return rtn;
202
203found_entry:
204 /* We found a matching entry which is either in progress or done. */
205 age = jiffies - rp->c_timestamp;
206 rp->c_timestamp = jiffies;
207 lru_put_end(rp);
208
209 rtn = RC_DROPIT;
210 /* Request being processed or excessive rexmits */
211 if (rp->c_state == RC_INPROG || age < RC_DELAY)
212 goto out;
213
214 /* From the hall of fame of impractical attacks:
215 * Is this a user who tries to snoop on the cache? */
216 rtn = RC_DOIT;
217 if (!rqstp->rq_secure && rp->c_secure)
218 goto out;
219
220 /* Compose RPC reply header */
221 switch (rp->c_type) {
222 case RC_NOCACHE:
223 break;
224 case RC_REPLSTAT:
225 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
226 rtn = RC_REPLY;
227 break;
228 case RC_REPLBUFF:
229 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
230 goto out; /* should not happen */
231 rtn = RC_REPLY;
232 break;
233 default:
234 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
235 rp->c_state = RC_UNUSED;
236 }
237
238 goto out;
239}
240
241/*
242 * Update a cache entry. This is called from nfsd_dispatch when
243 * the procedure has been executed and the complete reply is in
244 * rqstp->rq_res.
245 *
246 * We're copying around data here rather than swapping buffers because
247 * the toplevel loop requires max-sized buffers, which would be a waste
248 * of memory for a cache with a max reply size of 100 bytes (diropokres).
249 *
250 * If we should start to use different types of cache entries tailored
251 * specifically for attrstat and fh's, we may save even more space.
252 *
253 * Also note that a cachetype of RC_NOCACHE can legally be passed when
254 * nfsd failed to encode a reply that otherwise would have been cached.
255 * In this case, nfsd_cache_update is called with statp == NULL.
256 */
257void
Al Viroc7afef12006-10-19 23:29:02 -0700258nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700259{
260 struct svc_cacherep *rp;
261 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
262 int len;
263
264 if (!(rp = rqstp->rq_cacherep) || cache_disabled)
265 return;
266
267 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
268 len >>= 2;
Greg Banksfca42172009-04-01 07:28:13 +1100269
Linus Torvalds1da177e2005-04-16 15:20:36 -0700270 /* Don't cache excessive amounts of data and XDR failures */
271 if (!statp || len > (256 >> 2)) {
272 rp->c_state = RC_UNUSED;
273 return;
274 }
275
276 switch (cachetype) {
277 case RC_REPLSTAT:
278 if (len != 1)
279 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
280 rp->c_replstat = *statp;
281 break;
282 case RC_REPLBUFF:
283 cachv = &rp->c_replvec;
284 cachv->iov_base = kmalloc(len << 2, GFP_KERNEL);
285 if (!cachv->iov_base) {
286 spin_lock(&cache_lock);
287 rp->c_state = RC_UNUSED;
288 spin_unlock(&cache_lock);
289 return;
290 }
291 cachv->iov_len = len << 2;
292 memcpy(cachv->iov_base, statp, len << 2);
293 break;
294 }
295 spin_lock(&cache_lock);
296 lru_put_end(rp);
297 rp->c_secure = rqstp->rq_secure;
298 rp->c_type = cachetype;
299 rp->c_state = RC_DONE;
300 rp->c_timestamp = jiffies;
301 spin_unlock(&cache_lock);
302 return;
303}
304
305/*
306 * Copy cached reply to current reply buffer. Should always fit.
307 * FIXME as reply is in a page, we should just attach the page, and
308 * keep a refcount....
309 */
310static int
311nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
312{
313 struct kvec *vec = &rqstp->rq_res.head[0];
314
315 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
316 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
317 data->iov_len);
318 return 0;
319 }
320 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
321 vec->iov_len += data->iov_len;
322 return 1;
323}