2 * Request reply cache. This is currently a global cache, but this may
3 * change in the future and be a per-client cache.
5 * This code is heavily inspired by the 44BSD implementation, although
6 * it does things a bit differently.
8 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
14 /* Size of reply cache. Common values are:
20 #define CACHESIZE 1024
23 static struct hlist_head * cache_hash;
24 static struct list_head lru_head;
25 static int cache_disabled = 1;
28 * Calculate the hash index from an XID.
30 static inline u32 request_hash(u32 xid)
34 return h & (HASHSIZE-1);
37 static int nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
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.
44 static DEFINE_SPINLOCK(cache_lock);
46 int nfsd_reply_cache_init(void)
48 struct svc_cacherep *rp;
51 INIT_LIST_HEAD(&lru_head);
54 rp = kmalloc(sizeof(*rp), GFP_KERNEL);
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);
64 cache_hash = kcalloc (HASHSIZE, sizeof(struct hlist_head), GFP_KERNEL);
71 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
72 nfsd_reply_cache_shutdown();
76 void nfsd_reply_cache_shutdown(void)
78 struct svc_cacherep *rp;
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);
95 * Move cache entry to end of LRU list
98 lru_put_end(struct svc_cacherep *rp)
100 list_move_tail(&rp->c_lru, &lru_head);
104 * Move a cache entry from one hash list to another
107 hash_refile(struct svc_cacherep *rp)
109 hlist_del_init(&rp->c_hash);
110 hlist_add_head(&rp->c_hash, cache_hash + request_hash(rp->c_xid));
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.
119 nfsd_cache_lookup(struct svc_rqst *rqstp, int type)
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;
131 rqstp->rq_cacherep = NULL;
132 if (cache_disabled || type == RC_NOCACHE) {
133 nfsdstats.rcnocache++;
137 spin_lock(&cache_lock);
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) {
151 nfsdstats.rcmisses++;
153 /* This loop shouldn't take more than a few iterations normally */
156 list_for_each_entry(rp, &lru_head, c_lru) {
157 if (rp->c_state != RC_INPROG)
159 if (safe++ > CACHESIZE) {
160 printk("nfsd: loop in repcache LRU list\n");
167 /* All entries on the LRU are in-progress. This should not happen */
168 if (&rp->c_lru == &lru_head) {
169 static int complaints;
171 printk(KERN_WARNING "nfsd: all repcache entries locked!\n");
172 if (++complaints > 5) {
173 printk(KERN_WARNING "nfsd: disabling repcache.\n");
179 rqstp->rq_cacherep = rp;
180 rp->c_state = RC_INPROG;
183 memcpy(&rp->c_addr, svc_addr_in(rqstp), sizeof(rp->c_addr));
186 rp->c_timestamp = jiffies;
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;
195 rp->c_type = RC_NOCACHE;
197 spin_unlock(&cache_lock);
201 /* We found a matching entry which is either in progress or done. */
202 age = jiffies - rp->c_timestamp;
203 rp->c_timestamp = jiffies;
207 /* Request being processed or excessive rexmits */
208 if (rp->c_state == RC_INPROG || age < RC_DELAY)
211 /* From the hall of fame of impractical attacks:
212 * Is this a user who tries to snoop on the cache? */
214 if (!rqstp->rq_secure && rp->c_secure)
217 /* Compose RPC reply header */
218 switch (rp->c_type) {
222 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
226 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
227 goto out; /* should not happen */
231 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
232 rp->c_state = RC_UNUSED;
239 * Update a cache entry. This is called from nfsd_dispatch when
240 * the procedure has been executed and the complete reply is in
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).
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.
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.
255 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
257 struct svc_cacherep *rp;
258 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
261 if (!(rp = rqstp->rq_cacherep) || cache_disabled)
264 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
267 /* Don't cache excessive amounts of data and XDR failures */
268 if (!statp || len > (256 >> 2)) {
269 rp->c_state = RC_UNUSED;
276 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
277 rp->c_replstat = *statp;
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);
288 cachv->iov_len = len << 2;
289 memcpy(cachv->iov_base, statp, len << 2);
292 spin_lock(&cache_lock);
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);
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....
308 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
310 struct kvec *vec = &rqstp->rq_res.head[0];
312 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
313 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
317 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
318 vec->iov_len += data->iov_len;