xref: /linux/fs/nfsd/nfscache.c (revision 14b42963f64b98ab61fa9723c03d71aa5ef4f862)
1 /*
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
32 #define REQHASH(xid)		((((xid) >> 24) ^ (xid)) & (HASHSIZE-1))
33 
34 static struct hlist_head *	hash_list;
35 static struct list_head 	lru_head;
36 static int			cache_disabled = 1;
37 
38 static 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  */
45 static DEFINE_SPINLOCK(cache_lock);
46 
47 void
48 nfsd_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 
69 	hash_list = kmalloc (HASHSIZE * sizeof(struct hlist_head), GFP_KERNEL);
70 	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 	}
76 	memset(hash_list, 0, HASHSIZE * sizeof(struct hlist_head));
77 
78 	cache_disabled = 0;
79 }
80 
81 void
82 nfsd_cache_shutdown(void)
83 {
84 	struct svc_cacherep	*rp;
85 
86 	while (!list_empty(&lru_head)) {
87 		rp = list_entry(lru_head.next, struct svc_cacherep, c_lru);
88 		if (rp->c_state == RC_DONE && rp->c_type == RC_REPLBUFF)
89 			kfree(rp->c_replvec.iov_base);
90 		list_del(&rp->c_lru);
91 		kfree(rp);
92 	}
93 
94 	cache_disabled = 1;
95 
96 	kfree (hash_list);
97 	hash_list = NULL;
98 }
99 
100 /*
101  * Move cache entry to end of LRU list
102  */
103 static void
104 lru_put_end(struct svc_cacherep *rp)
105 {
106 	list_move_tail(&rp->c_lru, &lru_head);
107 }
108 
109 /*
110  * Move a cache entry from one hash list to another
111  */
112 static void
113 hash_refile(struct svc_cacherep *rp)
114 {
115 	hlist_del_init(&rp->c_hash);
116 	hlist_add_head(&rp->c_hash, hash_list + REQHASH(rp->c_xid));
117 }
118 
119 /*
120  * Try to find an entry matching the current call in the cache. When none
121  * is found, we grab the oldest unlocked entry off the LRU list.
122  * Note that no operation within the loop may sleep.
123  */
124 int
125 nfsd_cache_lookup(struct svc_rqst *rqstp, int type)
126 {
127 	struct hlist_node	*hn;
128 	struct hlist_head 	*rh;
129 	struct svc_cacherep	*rp;
130 	u32			xid = rqstp->rq_xid,
131 				proto =  rqstp->rq_prot,
132 				vers = rqstp->rq_vers,
133 				proc = rqstp->rq_proc;
134 	unsigned long		age;
135 	int rtn;
136 
137 	rqstp->rq_cacherep = NULL;
138 	if (cache_disabled || type == RC_NOCACHE) {
139 		nfsdstats.rcnocache++;
140 		return RC_DOIT;
141 	}
142 
143 	spin_lock(&cache_lock);
144 	rtn = RC_DOIT;
145 
146 	rh = &hash_list[REQHASH(xid)];
147 	hlist_for_each_entry(rp, hn, rh, c_hash) {
148 		if (rp->c_state != RC_UNUSED &&
149 		    xid == rp->c_xid && proc == rp->c_proc &&
150 		    proto == rp->c_prot && vers == rp->c_vers &&
151 		    time_before(jiffies, rp->c_timestamp + 120*HZ) &&
152 		    memcmp((char*)&rqstp->rq_addr, (char*)&rp->c_addr, sizeof(rp->c_addr))==0) {
153 			nfsdstats.rchits++;
154 			goto found_entry;
155 		}
156 	}
157 	nfsdstats.rcmisses++;
158 
159 	/* This loop shouldn't take more than a few iterations normally */
160 	{
161 	int	safe = 0;
162 	list_for_each_entry(rp, &lru_head, c_lru) {
163 		if (rp->c_state != RC_INPROG)
164 			break;
165 		if (safe++ > CACHESIZE) {
166 			printk("nfsd: loop in repcache LRU list\n");
167 			cache_disabled = 1;
168 			goto out;
169 		}
170 	}
171 	}
172 
173 	/* This should not happen */
174 	if (rp == NULL) {
175 		static int	complaints;
176 
177 		printk(KERN_WARNING "nfsd: all repcache entries locked!\n");
178 		if (++complaints > 5) {
179 			printk(KERN_WARNING "nfsd: disabling repcache.\n");
180 			cache_disabled = 1;
181 		}
182 		goto out;
183 	}
184 
185 	rqstp->rq_cacherep = rp;
186 	rp->c_state = RC_INPROG;
187 	rp->c_xid = xid;
188 	rp->c_proc = proc;
189 	rp->c_addr = rqstp->rq_addr;
190 	rp->c_prot = proto;
191 	rp->c_vers = vers;
192 	rp->c_timestamp = jiffies;
193 
194 	hash_refile(rp);
195 
196 	/* release any buffer */
197 	if (rp->c_type == RC_REPLBUFF) {
198 		kfree(rp->c_replvec.iov_base);
199 		rp->c_replvec.iov_base = NULL;
200 	}
201 	rp->c_type = RC_NOCACHE;
202  out:
203 	spin_unlock(&cache_lock);
204 	return rtn;
205 
206 found_entry:
207 	/* We found a matching entry which is either in progress or done. */
208 	age = jiffies - rp->c_timestamp;
209 	rp->c_timestamp = jiffies;
210 	lru_put_end(rp);
211 
212 	rtn = RC_DROPIT;
213 	/* Request being processed or excessive rexmits */
214 	if (rp->c_state == RC_INPROG || age < RC_DELAY)
215 		goto out;
216 
217 	/* From the hall of fame of impractical attacks:
218 	 * Is this a user who tries to snoop on the cache? */
219 	rtn = RC_DOIT;
220 	if (!rqstp->rq_secure && rp->c_secure)
221 		goto out;
222 
223 	/* Compose RPC reply header */
224 	switch (rp->c_type) {
225 	case RC_NOCACHE:
226 		break;
227 	case RC_REPLSTAT:
228 		svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
229 		rtn = RC_REPLY;
230 		break;
231 	case RC_REPLBUFF:
232 		if (!nfsd_cache_append(rqstp, &rp->c_replvec))
233 			goto out;	/* should not happen */
234 		rtn = RC_REPLY;
235 		break;
236 	default:
237 		printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
238 		rp->c_state = RC_UNUSED;
239 	}
240 
241 	goto out;
242 }
243 
244 /*
245  * Update a cache entry. This is called from nfsd_dispatch when
246  * the procedure has been executed and the complete reply is in
247  * rqstp->rq_res.
248  *
249  * We're copying around data here rather than swapping buffers because
250  * the toplevel loop requires max-sized buffers, which would be a waste
251  * of memory for a cache with a max reply size of 100 bytes (diropokres).
252  *
253  * If we should start to use different types of cache entries tailored
254  * specifically for attrstat and fh's, we may save even more space.
255  *
256  * Also note that a cachetype of RC_NOCACHE can legally be passed when
257  * nfsd failed to encode a reply that otherwise would have been cached.
258  * In this case, nfsd_cache_update is called with statp == NULL.
259  */
260 void
261 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, u32 *statp)
262 {
263 	struct svc_cacherep *rp;
264 	struct kvec	*resv = &rqstp->rq_res.head[0], *cachv;
265 	int		len;
266 
267 	if (!(rp = rqstp->rq_cacherep) || cache_disabled)
268 		return;
269 
270 	len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
271 	len >>= 2;
272 
273 	/* Don't cache excessive amounts of data and XDR failures */
274 	if (!statp || len > (256 >> 2)) {
275 		rp->c_state = RC_UNUSED;
276 		return;
277 	}
278 
279 	switch (cachetype) {
280 	case RC_REPLSTAT:
281 		if (len != 1)
282 			printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
283 		rp->c_replstat = *statp;
284 		break;
285 	case RC_REPLBUFF:
286 		cachv = &rp->c_replvec;
287 		cachv->iov_base = kmalloc(len << 2, GFP_KERNEL);
288 		if (!cachv->iov_base) {
289 			spin_lock(&cache_lock);
290 			rp->c_state = RC_UNUSED;
291 			spin_unlock(&cache_lock);
292 			return;
293 		}
294 		cachv->iov_len = len << 2;
295 		memcpy(cachv->iov_base, statp, len << 2);
296 		break;
297 	}
298 	spin_lock(&cache_lock);
299 	lru_put_end(rp);
300 	rp->c_secure = rqstp->rq_secure;
301 	rp->c_type = cachetype;
302 	rp->c_state = RC_DONE;
303 	rp->c_timestamp = jiffies;
304 	spin_unlock(&cache_lock);
305 	return;
306 }
307 
308 /*
309  * Copy cached reply to current reply buffer. Should always fit.
310  * FIXME as reply is in a page, we should just attach the page, and
311  * keep a refcount....
312  */
313 static int
314 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
315 {
316 	struct kvec	*vec = &rqstp->rq_res.head[0];
317 
318 	if (vec->iov_len + data->iov_len > PAGE_SIZE) {
319 		printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
320 				data->iov_len);
321 		return 0;
322 	}
323 	memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
324 	vec->iov_len += data->iov_len;
325 	return 1;
326 }
327