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)            ((((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         if (hash_list)
  97                 kfree (hash_list);
  98         hash_list = NULL;
  99 }
 100
 101 /*
 102  * Move cache entry to end of LRU list
 103  */
 104 static void
 105 lru_put_end(struct svc_cacherep *rp)
 106 {
 107         list_del(&rp->c_lru);
 108         list_add_tail(&rp->c_lru, &lru_head);
 109 }
 110
 111 /*
 112  * Move a cache entry from one hash list to another
 113  */
 114 static void
 115 hash_refile(struct svc_cacherep *rp)
 116 {
 117         hlist_del_init(&rp->c_hash);
 118         hlist_add_head(&rp->c_hash, hash_list + REQHASH(rp->c_xid));
 119 }
 120
 121 /*
 122  * Try to find an entry matching the current call in the cache. When none
 123  * is found, we grab the oldest unlocked entry off the LRU list.
 124  * Note that no operation within the loop may sleep.
 125  */
 126 int
 127 nfsd_cache_lookup(struct svc_rqst *rqstp, int type)
 128 {
 129         struct hlist_node       *hn;
 130         struct hlist_head       *rh;
 131         struct svc_cacherep     *rp;
 132         u32                     xid = rqstp->rq_xid,
 133                                 proto =  rqstp->rq_prot,
 134                                 vers = rqstp->rq_vers,
 135                                 proc = rqstp->rq_proc;
 136         unsigned long           age;
 137         int rtn;
 138
 139         rqstp->rq_cacherep = NULL;
 140         if (cache_disabled || type == RC_NOCACHE) {
 141                 nfsdstats.rcnocache++;
 142                 return RC_DOIT;
 143         }
 144
 145         spin_lock(&cache_lock);
 146         rtn = RC_DOIT;
 147
 148         rh = &hash_list[REQHASH(xid)];
 149         hlist_for_each_entry(rp, hn, rh, c_hash) {
 150                 if (rp->c_state != RC_UNUSED &&
 151                     xid == rp->c_xid && proc == rp->c_proc &&
 152                     proto == rp->c_prot && vers == rp->c_vers &&
 153                     time_before(jiffies, rp->c_timestamp + 120*HZ) &&
 154                     memcmp((char*)&rqstp->rq_addr, (char*)&rp->c_addr, sizeof(rp->c_addr))==0) {
 155                         nfsdstats.rchits++;
 156                         goto found_entry;
 157                 }
 158         }
 159         nfsdstats.rcmisses++;
 160
 161         /* This loop shouldn't take more than a few iterations normally */
 162         {
 163         int     safe = 0;
 164         list_for_each_entry(rp, &lru_head, c_lru) {
 165                 if (rp->c_state != RC_INPROG)
 166                         break;
 167                 if (safe++ > CACHESIZE) {
 168                         printk("nfsd: loop in repcache LRU list\n");
 169                         cache_disabled = 1;
 170                         goto out;
 171                 }
 172         }
 173         }
 174
 175         /* This should not happen */
 176         if (rp == NULL) {
 177                 static int      complaints;
 178
 179                 printk(KERN_WARNING "nfsd: all repcache entries locked!\n");
 180                 if (++complaints > 5) {
 181                         printk(KERN_WARNING "nfsd: disabling repcache.\n");
 182                         cache_disabled = 1;
 183                 }
 184                 goto out;
 185         }
 186
 187         rqstp->rq_cacherep = rp;
 188         rp->c_state = RC_INPROG;
 189         rp->c_xid = xid;
 190         rp->c_proc = proc;
 191         rp->c_addr = rqstp->rq_addr;
 192         rp->c_prot = proto;
 193         rp->c_vers = vers;
 194         rp->c_timestamp = jiffies;
 195
 196         hash_refile(rp);
 197
 198         /* release any buffer */
 199         if (rp->c_type == RC_REPLBUFF) {
 200                 kfree(rp->c_replvec.iov_base);
 201                 rp->c_replvec.iov_base = NULL;
 202         }
 203         rp->c_type = RC_NOCACHE;
 204  out:
 205         spin_unlock(&cache_lock);
 206         return rtn;
 207
 208 found_entry:
 209         /* We found a matching entry which is either in progress or done. */
 210         age = jiffies - rp->c_timestamp;
 211         rp->c_timestamp = jiffies;
 212         lru_put_end(rp);
 213
 214         rtn = RC_DROPIT;
 215         /* Request being processed or excessive rexmits */
 216         if (rp->c_state == RC_INPROG || age < RC_DELAY)
 217                 goto out;
 218
 219         /* From the hall of fame of impractical attacks:
 220          * Is this a user who tries to snoop on the cache? */
 221         rtn = RC_DOIT;
 222         if (!rqstp->rq_secure && rp->c_secure)
 223                 goto out;
 224
 225         /* Compose RPC reply header */
 226         switch (rp->c_type) {
 227         case RC_NOCACHE:
 228                 break;
 229         case RC_REPLSTAT:
 230                 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
 231                 rtn = RC_REPLY;
 232                 break;
 233         case RC_REPLBUFF:
 234                 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
 235                         goto out;       /* should not happen */
 236                 rtn = RC_REPLY;
 237                 break;
 238         default:
 239                 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
 240                 rp->c_state = RC_UNUSED;
 241         }
 242
 243         goto out;
 244 }
 245
 246 /*
 247  * Update a cache entry. This is called from nfsd_dispatch when
 248  * the procedure has been executed and the complete reply is in
 249  * rqstp->rq_res.
 250  *
 251  * We're copying around data here rather than swapping buffers because
 252  * the toplevel loop requires max-sized buffers, which would be a waste
 253  * of memory for a cache with a max reply size of 100 bytes (diropokres).
 254  *
 255  * If we should start to use different types of cache entries tailored
 256  * specifically for attrstat and fh's, we may save even more space.
 257  *
 258  * Also note that a cachetype of RC_NOCACHE can legally be passed when
 259  * nfsd failed to encode a reply that otherwise would have been cached.
 260  * In this case, nfsd_cache_update is called with statp == NULL.
 261  */
 262 void
 263 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, u32 *statp)
 264 {
 265         struct svc_cacherep *rp;
 266         struct kvec     *resv = &rqstp->rq_res.head[0], *cachv;
 267         int             len;
 268
 269         if (!(rp = rqstp->rq_cacherep) || cache_disabled)
 270                 return;
 271
 272         len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
 273         len >>= 2;
 274
 275         /* Don't cache excessive amounts of data and XDR failures */
 276         if (!statp || len > (256 >> 2)) {
 277                 rp->c_state = RC_UNUSED;
 278                 return;
 279         }
 280
 281         switch (cachetype) {
 282         case RC_REPLSTAT:
 283                 if (len != 1)
 284                         printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
 285                 rp->c_replstat = *statp;
 286                 break;
 287         case RC_REPLBUFF:
 288                 cachv = &rp->c_replvec;
 289                 cachv->iov_base = kmalloc(len << 2, GFP_KERNEL);
 290                 if (!cachv->iov_base) {
 291                         spin_lock(&cache_lock);
 292                         rp->c_state = RC_UNUSED;
 293                         spin_unlock(&cache_lock);
 294                         return;
 295                 }
 296                 cachv->iov_len = len << 2;
 297                 memcpy(cachv->iov_base, statp, len << 2);
 298                 break;
 299         }
 300         spin_lock(&cache_lock);
 301         lru_put_end(rp);
 302         rp->c_secure = rqstp->rq_secure;
 303         rp->c_type = cachetype;
 304         rp->c_state = RC_DONE;
 305         rp->c_timestamp = jiffies;
 306         spin_unlock(&cache_lock);
 307         return;
 308 }
 309
 310 /*
 311  * Copy cached reply to current reply buffer. Should always fit.
 312  * FIXME as reply is in a page, we should just attach the page, and
 313  * keep a refcount....
 314  */
 315 static int
 316 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
 317 {
 318         struct kvec     *vec = &rqstp->rq_res.head[0];
 319
 320         if (vec->iov_len + data->iov_len > PAGE_SIZE) {
 321                 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
 322                                 data->iov_len);
 323                 return 0;
 324         }
 325         memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
 326         vec->iov_len += data->iov_len;
 327         return 1;
 328 }