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/sched.h>
  15 #include <linux/malloc.h>
  16 #include <linux/string.h>
  17
  18 #include <linux/sunrpc/svc.h>
  19 #include <linux/nfsd/nfsd.h>
  20 #include <linux/nfsd/cache.h>
  21
  22 /* Size of reply cache. Common values are:
  23  * 4.3BSD:      128
  24  * 4.4BSD:      256
  25  * Solaris2:    1024
  26  * DEC Unix:    512-4096
  27  */
  28 #define CACHESIZE               1024
  29 #define HASHSIZE                64
  30 #define REQHASH(xid)            ((((xid) >> 24) ^ (xid)) & (HASHSIZE-1))
  31
  32 struct nfscache_head {
  33         struct svc_cacherep *   next;
  34         struct svc_cacherep *   prev;
  35 };
  36
  37 static struct nfscache_head *   hash_list;
  38 static struct svc_cacherep *    lru_head;
  39 static struct svc_cacherep *    lru_tail;
  40 static struct svc_cacherep *    nfscache;
  41 static int                      cache_initialized;
  42 static int                      cache_disabled = 1;
  43
  44 static int      nfsd_cache_append(struct svc_rqst *rqstp, struct svc_buf *data);
  45
  46 void
  47 nfsd_cache_init(void)
  48 {
  49         struct svc_cacherep     *rp;
  50         struct nfscache_head    *rh;
  51         size_t                  i;
  52         unsigned long           order;
  53
  54         if (cache_initialized)
  55                 return;
  56
  57         i = CACHESIZE * sizeof (struct svc_cacherep);
  58         for (order = 0; (PAGE_SIZE << order) < i; order++)
  59                 ;
  60         nfscache = (struct svc_cacherep *)
  61                 __get_free_pages(GFP_KERNEL, order);
  62         if (!nfscache) {
  63                 printk (KERN_ERR "nfsd: cannot allocate %Zd bytes for reply cache\n", i);
  64                 return;
  65         }
  66         memset(nfscache, 0, i);
  67
  68         i = HASHSIZE * sizeof (struct nfscache_head);
  69         hash_list = kmalloc (i, GFP_KERNEL);
  70         if (!hash_list) {
  71                 free_pages ((unsigned long)nfscache, order);
  72                 nfscache = NULL;
  73                 printk (KERN_ERR "nfsd: cannot allocate %Zd bytes for hash list\n", i);
  74                 return;
  75         }
  76
  77         for (i = 0, rh = hash_list; i < HASHSIZE; i++, rh++)
  78                 rh->next = rh->prev = (struct svc_cacherep *) rh;
  79
  80         for (i = 0, rp = nfscache; i < CACHESIZE; i++, rp++) {
  81                 rp->c_state = RC_UNUSED;
  82                 rp->c_type = RC_NOCACHE;
  83                 rp->c_hash_next =
  84                 rp->c_hash_prev = rp;
  85                 rp->c_lru_next = rp + 1;
  86                 rp->c_lru_prev = rp - 1;
  87         }
  88         lru_head = nfscache;
  89         lru_tail = nfscache + CACHESIZE - 1;
  90         lru_head->c_lru_prev = NULL;
  91         lru_tail->c_lru_next = NULL;
  92
  93         cache_initialized = 1;
  94         cache_disabled = 0;
  95 }
  96
  97 void
  98 nfsd_cache_shutdown(void)
  99 {
 100         struct svc_cacherep     *rp;
 101         size_t                  i;
 102         unsigned long           order;
 103
 104         if (!cache_initialized)
 105                 return;
 106
 107         for (rp = lru_head; rp; rp = rp->c_lru_next) {
 108                 if (rp->c_state == RC_DONE && rp->c_type == RC_REPLBUFF)
 109                         kfree(rp->c_replbuf.buf);
 110         }
 111
 112         cache_initialized = 0;
 113         cache_disabled = 1;
 114
 115         i = CACHESIZE * sizeof (struct svc_cacherep);
 116         for (order = 0; (PAGE_SIZE << order) < i; order++)
 117                 ;
 118         free_pages ((unsigned long)nfscache, order);
 119         nfscache = NULL;
 120         kfree (hash_list);
 121         hash_list = NULL;
 122 }
 123
 124 /*
 125  * Move cache entry to front of LRU list
 126  */
 127 static void
 128 lru_put_front(struct svc_cacherep *rp)
 129 {
 130         struct svc_cacherep     *prev = rp->c_lru_prev,
 131                                 *next = rp->c_lru_next;
 132
 133         if (prev)
 134                 prev->c_lru_next = next;
 135         else
 136                 lru_head = next;
 137         if (next)
 138                 next->c_lru_prev = prev;
 139         else
 140                 lru_tail = prev;
 141
 142         rp->c_lru_next = lru_head;
 143         rp->c_lru_prev = NULL;
 144         if (lru_head)
 145                 lru_head->c_lru_prev = rp;
 146         lru_head = rp;
 147 }
 148
 149 /*
 150  * Move a cache entry from one hash list to another
 151  */
 152 static void
 153 hash_refile(struct svc_cacherep *rp)
 154 {
 155         struct svc_cacherep     *prev = rp->c_hash_prev,
 156                                 *next = rp->c_hash_next;
 157         struct nfscache_head    *head = hash_list + REQHASH(rp->c_xid);
 158
 159         prev->c_hash_next = next;
 160         next->c_hash_prev = prev;
 161
 162         rp->c_hash_next = head->next;
 163         rp->c_hash_prev = (struct svc_cacherep *) head;
 164         head->next->c_hash_prev = rp;
 165         head->next = rp;
 166 }
 167
 168 /*
 169  * Try to find an entry matching the current call in the cache. When none
 170  * is found, we grab the oldest unlocked entry off the LRU list.
 171  * Note that no operation within the loop may sleep.
 172  */
 173 int
 174 nfsd_cache_lookup(struct svc_rqst *rqstp, int type)
 175 {
 176         struct svc_cacherep     *rh, *rp;
 177         u32                     xid = rqstp->rq_xid,
 178                                 proto =  rqstp->rq_prot,
 179                                 vers = rqstp->rq_vers,
 180                                 proc = rqstp->rq_proc;
 181         unsigned long           age;
 182
 183         rqstp->rq_cacherep = NULL;
 184         if (cache_disabled || type == RC_NOCACHE) {
 185                 nfsdstats.rcnocache++;
 186                 return RC_DOIT;
 187         }
 188
 189         rp = rh = (struct svc_cacherep *) &hash_list[REQHASH(xid)];
 190         while ((rp = rp->c_hash_next) != rh) {
 191                 if (rp->c_state != RC_UNUSED &&
 192                     xid == rp->c_xid && proc == rp->c_proc &&
 193                     proto == rp->c_prot && vers == rp->c_vers &&
 194                     time_before(jiffies, rp->c_timestamp + 120*HZ) &&
 195                     memcmp((char*)&rqstp->rq_addr, (char*)&rp->c_addr, rqstp->rq_addrlen)==0) {
 196                         nfsdstats.rchits++;
 197                         goto found_entry;
 198                 }
 199         }
 200         nfsdstats.rcmisses++;
 201
 202         /* This loop shouldn't take more than a few iterations normally */
 203         {
 204         int     safe = 0;
 205         for (rp = lru_tail; rp; rp = rp->c_lru_prev) {
 206                 if (rp->c_state != RC_INPROG)
 207                         break;
 208                 if (safe++ > CACHESIZE) {
 209                         printk("nfsd: loop in repcache LRU list\n");
 210                         cache_disabled = 1;
 211                         return RC_DOIT;
 212                 }
 213         }
 214         }
 215
 216         /* This should not happen */
 217         if (rp == NULL) {
 218                 static int      complaints;
 219
 220                 printk(KERN_WARNING "nfsd: all repcache entries locked!\n");
 221                 if (++complaints > 5) {
 222                         printk(KERN_WARNING "nfsd: disabling repcache.\n");
 223                         cache_disabled = 1;
 224                 }
 225                 return RC_DOIT;
 226         }
 227
 228         rqstp->rq_cacherep = rp;
 229         rp->c_state = RC_INPROG;
 230         rp->c_xid = xid;
 231         rp->c_proc = proc;
 232         rp->c_addr = rqstp->rq_addr;
 233         rp->c_prot = proto;
 234         rp->c_vers = vers;
 235         rp->c_timestamp = jiffies;
 236
 237         hash_refile(rp);
 238
 239         /* release any buffer */
 240         if (rp->c_type == RC_REPLBUFF) {
 241                 kfree(rp->c_replbuf.buf);
 242                 rp->c_replbuf.buf = NULL;
 243         }
 244         rp->c_type = RC_NOCACHE;
 245
 246         return RC_DOIT;
 247
 248 found_entry:
 249         /* We found a matching entry which is either in progress or done. */
 250         age = jiffies - rp->c_timestamp;
 251         rp->c_timestamp = jiffies;
 252         lru_put_front(rp);
 253
 254         /* Request being processed or excessive rexmits */
 255         if (rp->c_state == RC_INPROG || age < RC_DELAY)
 256                 return RC_DROPIT;
 257
 258         /* From the hall of fame of impractical attacks:
 259          * Is this a user who tries to snoop on the cache? */
 260         if (!rqstp->rq_secure && rp->c_secure)
 261                 return RC_DOIT;
 262
 263         /* Compose RPC reply header */
 264         switch (rp->c_type) {
 265         case RC_NOCACHE:
 266                 return RC_DOIT;
 267         case RC_REPLSTAT:
 268                 svc_putlong(&rqstp->rq_resbuf, rp->c_replstat);
 269                 break;
 270         case RC_REPLBUFF:
 271                 if (!nfsd_cache_append(rqstp, &rp->c_replbuf))
 272                         return RC_DOIT; /* should not happen */
 273                 break;
 274         default:
 275                 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
 276                 rp->c_state = RC_UNUSED;
 277                 return RC_DOIT;
 278         }
 279
 280         return RC_REPLY;
 281 }
 282
 283 /*
 284  * Update a cache entry. This is called from nfsd_dispatch when
 285  * the procedure has been executed and the complete reply is in
 286  * rqstp->rq_res.
 287  *
 288  * We're copying around data here rather than swapping buffers because
 289  * the toplevel loop requires max-sized buffers, which would be a waste
 290  * of memory for a cache with a max reply size of 100 bytes (diropokres).
 291  *
 292  * If we should start to use different types of cache entries tailored
 293  * specifically for attrstat and fh's, we may save even more space.
 294  *
 295  * Also note that a cachetype of RC_NOCACHE can legally be passed when
 296  * nfsd failed to encode a reply that otherwise would have been cached.
 297  * In this case, nfsd_cache_update is called with statp == NULL.
 298  */
 299 void
 300 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, u32 *statp)
 301 {
 302         struct svc_cacherep *rp;
 303         struct svc_buf  *resp = &rqstp->rq_resbuf, *cachp;
 304         int             len;
 305
 306         if (!(rp = rqstp->rq_cacherep) || cache_disabled)
 307                 return;
 308
 309         len = resp->len - (statp - resp->base);
 310
 311         /* Don't cache excessive amounts of data and XDR failures */
 312         if (!statp || len > (256 >> 2)) {
 313                 rp->c_state = RC_UNUSED;
 314                 return;
 315         }
 316
 317         switch (cachetype) {
 318         case RC_REPLSTAT:
 319                 if (len != 1)
 320                         printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
 321                 rp->c_replstat = *statp;
 322                 break;
 323         case RC_REPLBUFF:
 324                 cachp = &rp->c_replbuf;
 325                 cachp->buf = (u32 *) kmalloc(len << 2, GFP_KERNEL);
 326                 if (!cachp->buf) {
 327                         rp->c_state = RC_UNUSED;
 328                         return;
 329                 }
 330                 cachp->len = len;
 331                 memcpy(cachp->buf, statp, len << 2);
 332                 break;
 333         }
 334
 335         lru_put_front(rp);
 336         rp->c_secure = rqstp->rq_secure;
 337         rp->c_type = cachetype;
 338         rp->c_state = RC_DONE;
 339         rp->c_timestamp = jiffies;
 340
 341         return;
 342 }
 343
 344 /*
 345  * Copy cached reply to current reply buffer. Should always fit.
 346  */
 347 static int
 348 nfsd_cache_append(struct svc_rqst *rqstp, struct svc_buf *data)
 349 {
 350         struct svc_buf  *resp = &rqstp->rq_resbuf;
 351
 352         if (resp->len + data->len > resp->buflen) {
 353                 printk(KERN_WARNING "nfsd: cached reply too large (%d).\n",
 354                                 data->len);
 355                 return 0;
 356         }
 357         memcpy(resp->buf, data->buf, data->len << 2);
 358         resp->buf += data->len;
 359         resp->len += data->len;
 360         return 1;
 361 }