blob: 578f2c9d56bec09899d4dc1c0ca135320e7c6bf9 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/fs/nfsd/nfscache.c
3 *
4 * Request reply cache. This is currently a global cache, but this may
5 * change in the future and be a per-client cache.
6 *
7 * This code is heavily inspired by the 44BSD implementation, although
8 * it does things a bit differently.
9 *
10 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
11 */
12
13#include <linux/kernel.h>
14#include <linux/time.h>
15#include <linux/slab.h>
16#include <linux/string.h>
17#include <linux/spinlock.h>
18#include <linux/list.h>
19
20#include <linux/sunrpc/svc.h>
21#include <linux/nfsd/nfsd.h>
22#include <linux/nfsd/cache.h>
23
24/* Size of reply cache. Common values are:
25 * 4.3BSD: 128
26 * 4.4BSD: 256
27 * Solaris2: 1024
28 * DEC Unix: 512-4096
29 */
30#define CACHESIZE 1024
31#define HASHSIZE 64
Al Viroc7afef12006-10-19 23:29:02 -070032#define REQHASH(xid) (((((__force __u32)xid) >> 24) ^ ((__force __u32)xid)) & (HASHSIZE-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -070033
34static struct hlist_head * hash_list;
35static struct list_head lru_head;
36static int cache_disabled = 1;
37
38static int nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
39
40/*
41 * locking for the reply cache:
42 * A cache entry is "single use" if c_state == RC_INPROG
43 * Otherwise, it when accessing _prev or _next, the lock must be held.
44 */
45static DEFINE_SPINLOCK(cache_lock);
46
47void
48nfsd_cache_init(void)
49{
50 struct svc_cacherep *rp;
51 int i;
52
53 INIT_LIST_HEAD(&lru_head);
54 i = CACHESIZE;
55 while(i) {
56 rp = kmalloc(sizeof(*rp), GFP_KERNEL);
57 if (!rp) break;
58 list_add(&rp->c_lru, &lru_head);
59 rp->c_state = RC_UNUSED;
60 rp->c_type = RC_NOCACHE;
61 INIT_HLIST_NODE(&rp->c_hash);
62 i--;
63 }
64
65 if (i)
66 printk (KERN_ERR "nfsd: cannot allocate all %d cache entries, only got %d\n",
67 CACHESIZE, CACHESIZE-i);
68
Yan Burman4b3bb062006-12-08 02:39:41 -080069 hash_list = kcalloc (HASHSIZE, sizeof(struct hlist_head), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -070070 if (!hash_list) {
71 nfsd_cache_shutdown();
72 printk (KERN_ERR "nfsd: cannot allocate %Zd bytes for hash list\n",
73 HASHSIZE * sizeof(struct hlist_head));
74 return;
75 }
Linus Torvalds1da177e2005-04-16 15:20:36 -070076
77 cache_disabled = 0;
78}
79
80void
81nfsd_cache_shutdown(void)
82{
83 struct svc_cacherep *rp;
84
85 while (!list_empty(&lru_head)) {
86 rp = list_entry(lru_head.next, struct svc_cacherep, c_lru);
87 if (rp->c_state == RC_DONE && rp->c_type == RC_REPLBUFF)
88 kfree(rp->c_replvec.iov_base);
89 list_del(&rp->c_lru);
90 kfree(rp);
91 }
92
93 cache_disabled = 1;
94
Jesper Juhlf99d49a2005-11-07 01:01:34 -080095 kfree (hash_list);
Linus Torvalds1da177e2005-04-16 15:20:36 -070096 hash_list = NULL;
97}
98
99/*
100 * Move cache entry to end of LRU list
101 */
102static void
103lru_put_end(struct svc_cacherep *rp)
104{
Akinobu Mitaf1166292006-06-26 00:24:46 -0700105 list_move_tail(&rp->c_lru, &lru_head);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106}
107
108/*
109 * Move a cache entry from one hash list to another
110 */
111static void
112hash_refile(struct svc_cacherep *rp)
113{
114 hlist_del_init(&rp->c_hash);
115 hlist_add_head(&rp->c_hash, hash_list + REQHASH(rp->c_xid));
116}
117
118/*
119 * Try to find an entry matching the current call in the cache. When none
120 * is found, we grab the oldest unlocked entry off the LRU list.
121 * Note that no operation within the loop may sleep.
122 */
123int
124nfsd_cache_lookup(struct svc_rqst *rqstp, int type)
125{
126 struct hlist_node *hn;
127 struct hlist_head *rh;
128 struct svc_cacherep *rp;
Al Viroc7afef12006-10-19 23:29:02 -0700129 __be32 xid = rqstp->rq_xid;
130 u32 proto = rqstp->rq_prot,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131 vers = rqstp->rq_vers,
132 proc = rqstp->rq_proc;
133 unsigned long age;
134 int rtn;
135
136 rqstp->rq_cacherep = NULL;
137 if (cache_disabled || type == RC_NOCACHE) {
138 nfsdstats.rcnocache++;
139 return RC_DOIT;
140 }
141
142 spin_lock(&cache_lock);
143 rtn = RC_DOIT;
144
145 rh = &hash_list[REQHASH(xid)];
146 hlist_for_each_entry(rp, hn, rh, c_hash) {
147 if (rp->c_state != RC_UNUSED &&
148 xid == rp->c_xid && proc == rp->c_proc &&
149 proto == rp->c_prot && vers == rp->c_vers &&
150 time_before(jiffies, rp->c_timestamp + 120*HZ) &&
151 memcmp((char*)&rqstp->rq_addr, (char*)&rp->c_addr, sizeof(rp->c_addr))==0) {
152 nfsdstats.rchits++;
153 goto found_entry;
154 }
155 }
156 nfsdstats.rcmisses++;
157
158 /* This loop shouldn't take more than a few iterations normally */
159 {
160 int safe = 0;
161 list_for_each_entry(rp, &lru_head, c_lru) {
162 if (rp->c_state != RC_INPROG)
163 break;
164 if (safe++ > CACHESIZE) {
165 printk("nfsd: loop in repcache LRU list\n");
166 cache_disabled = 1;
167 goto out;
168 }
169 }
170 }
171
172 /* This should not happen */
173 if (rp == NULL) {
174 static int complaints;
175
176 printk(KERN_WARNING "nfsd: all repcache entries locked!\n");
177 if (++complaints > 5) {
178 printk(KERN_WARNING "nfsd: disabling repcache.\n");
179 cache_disabled = 1;
180 }
181 goto out;
182 }
183
184 rqstp->rq_cacherep = rp;
185 rp->c_state = RC_INPROG;
186 rp->c_xid = xid;
187 rp->c_proc = proc;
Chuck Lever27459f02007-02-12 00:53:34 -0800188 memcpy(&rp->c_addr, svc_addr_in(rqstp), sizeof(rp->c_addr));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189 rp->c_prot = proto;
190 rp->c_vers = vers;
191 rp->c_timestamp = jiffies;
192
193 hash_refile(rp);
194
195 /* release any buffer */
196 if (rp->c_type == RC_REPLBUFF) {
197 kfree(rp->c_replvec.iov_base);
198 rp->c_replvec.iov_base = NULL;
199 }
200 rp->c_type = RC_NOCACHE;
201 out:
202 spin_unlock(&cache_lock);
203 return rtn;
204
205found_entry:
206 /* We found a matching entry which is either in progress or done. */
207 age = jiffies - rp->c_timestamp;
208 rp->c_timestamp = jiffies;
209 lru_put_end(rp);
210
211 rtn = RC_DROPIT;
212 /* Request being processed or excessive rexmits */
213 if (rp->c_state == RC_INPROG || age < RC_DELAY)
214 goto out;
215
216 /* From the hall of fame of impractical attacks:
217 * Is this a user who tries to snoop on the cache? */
218 rtn = RC_DOIT;
219 if (!rqstp->rq_secure && rp->c_secure)
220 goto out;
221
222 /* Compose RPC reply header */
223 switch (rp->c_type) {
224 case RC_NOCACHE:
225 break;
226 case RC_REPLSTAT:
227 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
228 rtn = RC_REPLY;
229 break;
230 case RC_REPLBUFF:
231 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
232 goto out; /* should not happen */
233 rtn = RC_REPLY;
234 break;
235 default:
236 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
237 rp->c_state = RC_UNUSED;
238 }
239
240 goto out;
241}
242
243/*
244 * Update a cache entry. This is called from nfsd_dispatch when
245 * the procedure has been executed and the complete reply is in
246 * rqstp->rq_res.
247 *
248 * We're copying around data here rather than swapping buffers because
249 * the toplevel loop requires max-sized buffers, which would be a waste
250 * of memory for a cache with a max reply size of 100 bytes (diropokres).
251 *
252 * If we should start to use different types of cache entries tailored
253 * specifically for attrstat and fh's, we may save even more space.
254 *
255 * Also note that a cachetype of RC_NOCACHE can legally be passed when
256 * nfsd failed to encode a reply that otherwise would have been cached.
257 * In this case, nfsd_cache_update is called with statp == NULL.
258 */
259void
Al Viroc7afef12006-10-19 23:29:02 -0700260nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261{
262 struct svc_cacherep *rp;
263 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
264 int len;
265
266 if (!(rp = rqstp->rq_cacherep) || cache_disabled)
267 return;
268
269 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
270 len >>= 2;
271
272 /* Don't cache excessive amounts of data and XDR failures */
273 if (!statp || len > (256 >> 2)) {
274 rp->c_state = RC_UNUSED;
275 return;
276 }
277
278 switch (cachetype) {
279 case RC_REPLSTAT:
280 if (len != 1)
281 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
282 rp->c_replstat = *statp;
283 break;
284 case RC_REPLBUFF:
285 cachv = &rp->c_replvec;
286 cachv->iov_base = kmalloc(len << 2, GFP_KERNEL);
287 if (!cachv->iov_base) {
288 spin_lock(&cache_lock);
289 rp->c_state = RC_UNUSED;
290 spin_unlock(&cache_lock);
291 return;
292 }
293 cachv->iov_len = len << 2;
294 memcpy(cachv->iov_base, statp, len << 2);
295 break;
296 }
297 spin_lock(&cache_lock);
298 lru_put_end(rp);
299 rp->c_secure = rqstp->rq_secure;
300 rp->c_type = cachetype;
301 rp->c_state = RC_DONE;
302 rp->c_timestamp = jiffies;
303 spin_unlock(&cache_lock);
304 return;
305}
306
307/*
308 * Copy cached reply to current reply buffer. Should always fit.
309 * FIXME as reply is in a page, we should just attach the page, and
310 * keep a refcount....
311 */
312static int
313nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
314{
315 struct kvec *vec = &rqstp->rq_res.head[0];
316
317 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
318 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
319 data->iov_len);
320 return 0;
321 }
322 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
323 vec->iov_len += data->iov_len;
324 return 1;
325}