xref: /openbmc/linux/fs/nfsd/nfscache.c (revision a09d2831)
1 /*
2  * 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 
11 #include "nfsd.h"
12 #include "cache.h"
13 
14 /* Size of reply cache. Common values are:
15  * 4.3BSD:	128
16  * 4.4BSD:	256
17  * Solaris2:	1024
18  * DEC Unix:	512-4096
19  */
20 #define CACHESIZE		1024
21 #define HASHSIZE		64
22 
23 static struct hlist_head *	cache_hash;
24 static struct list_head 	lru_head;
25 static int			cache_disabled = 1;
26 
27 /*
28  * Calculate the hash index from an XID.
29  */
30 static inline u32 request_hash(u32 xid)
31 {
32 	u32 h = xid;
33 	h ^= (xid >> 24);
34 	return h & (HASHSIZE-1);
35 }
36 
37 static int	nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
38 
39 /*
40  * locking for the reply cache:
41  * A cache entry is "single use" if c_state == RC_INPROG
42  * Otherwise, it when accessing _prev or _next, the lock must be held.
43  */
44 static DEFINE_SPINLOCK(cache_lock);
45 
46 int nfsd_reply_cache_init(void)
47 {
48 	struct svc_cacherep	*rp;
49 	int			i;
50 
51 	INIT_LIST_HEAD(&lru_head);
52 	i = CACHESIZE;
53 	while (i) {
54 		rp = kmalloc(sizeof(*rp), GFP_KERNEL);
55 		if (!rp)
56 			goto out_nomem;
57 		list_add(&rp->c_lru, &lru_head);
58 		rp->c_state = RC_UNUSED;
59 		rp->c_type = RC_NOCACHE;
60 		INIT_HLIST_NODE(&rp->c_hash);
61 		i--;
62 	}
63 
64 	cache_hash = kcalloc (HASHSIZE, sizeof(struct hlist_head), GFP_KERNEL);
65 	if (!cache_hash)
66 		goto out_nomem;
67 
68 	cache_disabled = 0;
69 	return 0;
70 out_nomem:
71 	printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
72 	nfsd_reply_cache_shutdown();
73 	return -ENOMEM;
74 }
75 
76 void nfsd_reply_cache_shutdown(void)
77 {
78 	struct svc_cacherep	*rp;
79 
80 	while (!list_empty(&lru_head)) {
81 		rp = list_entry(lru_head.next, struct svc_cacherep, c_lru);
82 		if (rp->c_state == RC_DONE && rp->c_type == RC_REPLBUFF)
83 			kfree(rp->c_replvec.iov_base);
84 		list_del(&rp->c_lru);
85 		kfree(rp);
86 	}
87 
88 	cache_disabled = 1;
89 
90 	kfree (cache_hash);
91 	cache_hash = NULL;
92 }
93 
94 /*
95  * Move cache entry to end of LRU list
96  */
97 static void
98 lru_put_end(struct svc_cacherep *rp)
99 {
100 	list_move_tail(&rp->c_lru, &lru_head);
101 }
102 
103 /*
104  * Move a cache entry from one hash list to another
105  */
106 static void
107 hash_refile(struct svc_cacherep *rp)
108 {
109 	hlist_del_init(&rp->c_hash);
110 	hlist_add_head(&rp->c_hash, cache_hash + request_hash(rp->c_xid));
111 }
112 
113 /*
114  * Try to find an entry matching the current call in the cache. When none
115  * is found, we grab the oldest unlocked entry off the LRU list.
116  * Note that no operation within the loop may sleep.
117  */
118 int
119 nfsd_cache_lookup(struct svc_rqst *rqstp, int type)
120 {
121 	struct hlist_node	*hn;
122 	struct hlist_head 	*rh;
123 	struct svc_cacherep	*rp;
124 	__be32			xid = rqstp->rq_xid;
125 	u32			proto =  rqstp->rq_prot,
126 				vers = rqstp->rq_vers,
127 				proc = rqstp->rq_proc;
128 	unsigned long		age;
129 	int rtn;
130 
131 	rqstp->rq_cacherep = NULL;
132 	if (cache_disabled || type == RC_NOCACHE) {
133 		nfsdstats.rcnocache++;
134 		return RC_DOIT;
135 	}
136 
137 	spin_lock(&cache_lock);
138 	rtn = RC_DOIT;
139 
140 	rh = &cache_hash[request_hash(xid)];
141 	hlist_for_each_entry(rp, hn, rh, c_hash) {
142 		if (rp->c_state != RC_UNUSED &&
143 		    xid == rp->c_xid && proc == rp->c_proc &&
144 		    proto == rp->c_prot && vers == rp->c_vers &&
145 		    time_before(jiffies, rp->c_timestamp + 120*HZ) &&
146 		    memcmp((char*)&rqstp->rq_addr, (char*)&rp->c_addr, sizeof(rp->c_addr))==0) {
147 			nfsdstats.rchits++;
148 			goto found_entry;
149 		}
150 	}
151 	nfsdstats.rcmisses++;
152 
153 	/* This loop shouldn't take more than a few iterations normally */
154 	{
155 	int	safe = 0;
156 	list_for_each_entry(rp, &lru_head, c_lru) {
157 		if (rp->c_state != RC_INPROG)
158 			break;
159 		if (safe++ > CACHESIZE) {
160 			printk("nfsd: loop in repcache LRU list\n");
161 			cache_disabled = 1;
162 			goto out;
163 		}
164 	}
165 	}
166 
167 	/* All entries on the LRU are in-progress. This should not happen */
168 	if (&rp->c_lru == &lru_head) {
169 		static int	complaints;
170 
171 		printk(KERN_WARNING "nfsd: all repcache entries locked!\n");
172 		if (++complaints > 5) {
173 			printk(KERN_WARNING "nfsd: disabling repcache.\n");
174 			cache_disabled = 1;
175 		}
176 		goto out;
177 	}
178 
179 	rqstp->rq_cacherep = rp;
180 	rp->c_state = RC_INPROG;
181 	rp->c_xid = xid;
182 	rp->c_proc = proc;
183 	memcpy(&rp->c_addr, svc_addr_in(rqstp), sizeof(rp->c_addr));
184 	rp->c_prot = proto;
185 	rp->c_vers = vers;
186 	rp->c_timestamp = jiffies;
187 
188 	hash_refile(rp);
189 
190 	/* release any buffer */
191 	if (rp->c_type == RC_REPLBUFF) {
192 		kfree(rp->c_replvec.iov_base);
193 		rp->c_replvec.iov_base = NULL;
194 	}
195 	rp->c_type = RC_NOCACHE;
196  out:
197 	spin_unlock(&cache_lock);
198 	return rtn;
199 
200 found_entry:
201 	/* We found a matching entry which is either in progress or done. */
202 	age = jiffies - rp->c_timestamp;
203 	rp->c_timestamp = jiffies;
204 	lru_put_end(rp);
205 
206 	rtn = RC_DROPIT;
207 	/* Request being processed or excessive rexmits */
208 	if (rp->c_state == RC_INPROG || age < RC_DELAY)
209 		goto out;
210 
211 	/* From the hall of fame of impractical attacks:
212 	 * Is this a user who tries to snoop on the cache? */
213 	rtn = RC_DOIT;
214 	if (!rqstp->rq_secure && rp->c_secure)
215 		goto out;
216 
217 	/* Compose RPC reply header */
218 	switch (rp->c_type) {
219 	case RC_NOCACHE:
220 		break;
221 	case RC_REPLSTAT:
222 		svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
223 		rtn = RC_REPLY;
224 		break;
225 	case RC_REPLBUFF:
226 		if (!nfsd_cache_append(rqstp, &rp->c_replvec))
227 			goto out;	/* should not happen */
228 		rtn = RC_REPLY;
229 		break;
230 	default:
231 		printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
232 		rp->c_state = RC_UNUSED;
233 	}
234 
235 	goto out;
236 }
237 
238 /*
239  * Update a cache entry. This is called from nfsd_dispatch when
240  * the procedure has been executed and the complete reply is in
241  * rqstp->rq_res.
242  *
243  * We're copying around data here rather than swapping buffers because
244  * the toplevel loop requires max-sized buffers, which would be a waste
245  * of memory for a cache with a max reply size of 100 bytes (diropokres).
246  *
247  * If we should start to use different types of cache entries tailored
248  * specifically for attrstat and fh's, we may save even more space.
249  *
250  * Also note that a cachetype of RC_NOCACHE can legally be passed when
251  * nfsd failed to encode a reply that otherwise would have been cached.
252  * In this case, nfsd_cache_update is called with statp == NULL.
253  */
254 void
255 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
256 {
257 	struct svc_cacherep *rp;
258 	struct kvec	*resv = &rqstp->rq_res.head[0], *cachv;
259 	int		len;
260 
261 	if (!(rp = rqstp->rq_cacherep) || cache_disabled)
262 		return;
263 
264 	len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
265 	len >>= 2;
266 
267 	/* Don't cache excessive amounts of data and XDR failures */
268 	if (!statp || len > (256 >> 2)) {
269 		rp->c_state = RC_UNUSED;
270 		return;
271 	}
272 
273 	switch (cachetype) {
274 	case RC_REPLSTAT:
275 		if (len != 1)
276 			printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
277 		rp->c_replstat = *statp;
278 		break;
279 	case RC_REPLBUFF:
280 		cachv = &rp->c_replvec;
281 		cachv->iov_base = kmalloc(len << 2, GFP_KERNEL);
282 		if (!cachv->iov_base) {
283 			spin_lock(&cache_lock);
284 			rp->c_state = RC_UNUSED;
285 			spin_unlock(&cache_lock);
286 			return;
287 		}
288 		cachv->iov_len = len << 2;
289 		memcpy(cachv->iov_base, statp, len << 2);
290 		break;
291 	}
292 	spin_lock(&cache_lock);
293 	lru_put_end(rp);
294 	rp->c_secure = rqstp->rq_secure;
295 	rp->c_type = cachetype;
296 	rp->c_state = RC_DONE;
297 	rp->c_timestamp = jiffies;
298 	spin_unlock(&cache_lock);
299 	return;
300 }
301 
302 /*
303  * Copy cached reply to current reply buffer. Should always fit.
304  * FIXME as reply is in a page, we should just attach the page, and
305  * keep a refcount....
306  */
307 static int
308 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
309 {
310 	struct kvec	*vec = &rqstp->rq_res.head[0];
311 
312 	if (vec->iov_len + data->iov_len > PAGE_SIZE) {
313 		printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
314 				data->iov_len);
315 		return 0;
316 	}
317 	memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
318 	vec->iov_len += data->iov_len;
319 	return 1;
320 }
321