fs/nfsd/nfscache.c

   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)            (((((__force __u32)xid) >> 24) ^ ((__force __u32)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 int
  48 nfsd_reply_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)
  58                         goto out_nomem;
  59                 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
  66         hash_list = kcalloc (HASHSIZE, sizeof(struct hlist_head), GFP_KERNEL);
  67         if (!hash_list)
  68                 goto out_nomem;
  69
  70         cache_disabled = 0;
  71         return 0;
  72 out_nomem:
  73         printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
  74         nfsd_reply_cache_shutdown();
  75         return -ENOMEM;
  76 }
  77
  78 void
  79 nfsd_reply_cache_shutdown(void)
  80 {
  81         struct svc_cacherep     *rp;
  82
  83         while (!list_empty(&lru_head)) {
  84                 rp = list_entry(lru_head.next, struct svc_cacherep, c_lru);
  85                 if (rp->c_state == RC_DONE && rp->c_type == RC_REPLBUFF)
  86                         kfree(rp->c_replvec.iov_base);
  87                 list_del(&rp->c_lru);
  88                 kfree(rp);
  89         }
  90
  91         cache_disabled = 1;
  92
  93         kfree (hash_list);
  94         hash_list = NULL;
  95 }
  96
  97 /*
  98  * Move cache entry to end of LRU list
  99  */
 100 static void
 101 lru_put_end(struct svc_cacherep *rp)
 102 {
 103         list_move_tail(&rp->c_lru, &lru_head);
 104 }
 105
 106 /*
 107  * Move a cache entry from one hash list to another
 108  */
 109 static void
 110 hash_refile(struct svc_cacherep *rp)
 111 {
 112         hlist_del_init(&rp->c_hash);
 113         hlist_add_head(&rp->c_hash, hash_list + REQHASH(rp->c_xid));
 114 }
 115
 116 /*
 117  * Try to find an entry matching the current call in the cache. When none
 118  * is found, we grab the oldest unlocked entry off the LRU list.
 119  * Note that no operation within the loop may sleep.
 120  */
 121 int
 122 nfsd_cache_lookup(struct svc_rqst *rqstp, int type)
 123 {
 124         struct hlist_node       *hn;
 125         struct hlist_head       *rh;
 126         struct svc_cacherep     *rp;
 127         __be32                  xid = rqstp->rq_xid;
 128         u32                     proto =  rqstp->rq_prot,
 129                                 vers = rqstp->rq_vers,
 130                                 proc = rqstp->rq_proc;
 131         unsigned long           age;
 132         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
 143         rh = &hash_list[REQHASH(xid)];
 144         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
 170         /* This should not happen */
 171         if (rp == NULL) {
 172                 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;
 186         memcpy(&rp->c_addr, svc_addr_in(rqstp), sizeof(rp->c_addr));
 187         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
 203 found_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  */
 257 void
 258 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
 259 {
 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;
 269
 270         /* 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  */
 310 static int
 311 nfsd_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 }