fs/nfsd/nfscache.c

   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 <linux/slab.h>
  12 #include <linux/sunrpc/addr.h>
  13 #include <linux/highmem.h>
  14 #include <net/checksum.h>
  15
  16 #include "nfsd.h"
  17 #include "cache.h"
  18
  19 #define NFSDDBG_FACILITY        NFSDDBG_REPCACHE
  20
  21 #define HASHSIZE                64
  22
  23 static struct hlist_head *      cache_hash;
  24 static struct list_head         lru_head;
  25 static struct kmem_cache        *drc_slab;
  26 static unsigned int             num_drc_entries;
  27 static unsigned int             max_drc_entries;
  28
  29 /*
  30  * Calculate the hash index from an XID.
  31  */
  32 static inline u32 request_hash(u32 xid)
  33 {
  34         u32 h = xid;
  35         h ^= (xid >> 24);
  36         return h & (HASHSIZE-1);
  37 }
  38
  39 static int      nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
  40 static void     cache_cleaner_func(struct work_struct *unused);
  41 static int      nfsd_reply_cache_shrink(struct shrinker *shrink,
  42                                         struct shrink_control *sc);
  43
  44 struct shrinker nfsd_reply_cache_shrinker = {
  45         .shrink = nfsd_reply_cache_shrink,
  46         .seeks  = 1,
  47 };
  48
  49 /*
  50  * locking for the reply cache:
  51  * A cache entry is "single use" if c_state == RC_INPROG
  52  * Otherwise, it when accessing _prev or _next, the lock must be held.
  53  */
  54 static DEFINE_SPINLOCK(cache_lock);
  55 static DECLARE_DELAYED_WORK(cache_cleaner, cache_cleaner_func);
  56
  57 /*
  58  * Put a cap on the size of the DRC based on the amount of available
  59  * low memory in the machine.
  60  *
  61  *  64MB:    8192
  62  * 128MB:   11585
  63  * 256MB:   16384
  64  * 512MB:   23170
  65  *   1GB:   32768
  66  *   2GB:   46340
  67  *   4GB:   65536
  68  *   8GB:   92681
  69  *  16GB:  131072
  70  *
  71  * ...with a hard cap of 256k entries. In the worst case, each entry will be
  72  * ~1k, so the above numbers should give a rough max of the amount of memory
  73  * used in k.
  74  */
  75 static unsigned int
  76 nfsd_cache_size_limit(void)
  77 {
  78         unsigned int limit;
  79         unsigned long low_pages = totalram_pages - totalhigh_pages;
  80
  81         limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
  82         return min_t(unsigned int, limit, 256*1024);
  83 }
  84
  85 static struct svc_cacherep *
  86 nfsd_reply_cache_alloc(void)
  87 {
  88         struct svc_cacherep     *rp;
  89
  90         rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
  91         if (rp) {
  92                 rp->c_state = RC_UNUSED;
  93                 rp->c_type = RC_NOCACHE;
  94                 INIT_LIST_HEAD(&rp->c_lru);
  95                 INIT_HLIST_NODE(&rp->c_hash);
  96         }
  97         return rp;
  98 }
  99
 100 static void
 101 nfsd_reply_cache_free_locked(struct svc_cacherep *rp)
 102 {
 103         if (rp->c_type == RC_REPLBUFF)
 104                 kfree(rp->c_replvec.iov_base);
 105         hlist_del(&rp->c_hash);
 106         list_del(&rp->c_lru);
 107         --num_drc_entries;
 108         kmem_cache_free(drc_slab, rp);
 109 }
 110
 111 static void
 112 nfsd_reply_cache_free(struct svc_cacherep *rp)
 113 {
 114         spin_lock(&cache_lock);
 115         nfsd_reply_cache_free_locked(rp);
 116         spin_unlock(&cache_lock);
 117 }
 118
 119 int nfsd_reply_cache_init(void)
 120 {
 121         register_shrinker(&nfsd_reply_cache_shrinker);
 122         drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
 123                                         0, 0, NULL);
 124         if (!drc_slab)
 125                 goto out_nomem;
 126
 127         cache_hash = kcalloc(HASHSIZE, sizeof(struct hlist_head), GFP_KERNEL);
 128         if (!cache_hash)
 129                 goto out_nomem;
 130
 131         INIT_LIST_HEAD(&lru_head);
 132         max_drc_entries = nfsd_cache_size_limit();
 133         num_drc_entries = 0;
 134
 135         return 0;
 136 out_nomem:
 137         printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
 138         nfsd_reply_cache_shutdown();
 139         return -ENOMEM;
 140 }
 141
 142 void nfsd_reply_cache_shutdown(void)
 143 {
 144         struct svc_cacherep     *rp;
 145
 146         unregister_shrinker(&nfsd_reply_cache_shrinker);
 147         cancel_delayed_work_sync(&cache_cleaner);
 148
 149         while (!list_empty(&lru_head)) {
 150                 rp = list_entry(lru_head.next, struct svc_cacherep, c_lru);
 151                 nfsd_reply_cache_free_locked(rp);
 152         }
 153
 154         kfree (cache_hash);
 155         cache_hash = NULL;
 156
 157         if (drc_slab) {
 158                 kmem_cache_destroy(drc_slab);
 159                 drc_slab = NULL;
 160         }
 161 }
 162
 163 /*
 164  * Move cache entry to end of LRU list, and queue the cleaner to run if it's
 165  * not already scheduled.
 166  */
 167 static void
 168 lru_put_end(struct svc_cacherep *rp)
 169 {
 170         rp->c_timestamp = jiffies;
 171         list_move_tail(&rp->c_lru, &lru_head);
 172         schedule_delayed_work(&cache_cleaner, RC_EXPIRE);
 173 }
 174
 175 /*
 176  * Move a cache entry from one hash list to another
 177  */
 178 static void
 179 hash_refile(struct svc_cacherep *rp)
 180 {
 181         hlist_del_init(&rp->c_hash);
 182         hlist_add_head(&rp->c_hash, cache_hash + request_hash(rp->c_xid));
 183 }
 184
 185 static inline bool
 186 nfsd_cache_entry_expired(struct svc_cacherep *rp)
 187 {
 188         return rp->c_state != RC_INPROG &&
 189                time_after(jiffies, rp->c_timestamp + RC_EXPIRE);
 190 }
 191
 192 /*
 193  * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
 194  * Also prune the oldest ones when the total exceeds the max number of entries.
 195  */
 196 static void
 197 prune_cache_entries(void)
 198 {
 199         struct svc_cacherep *rp, *tmp;
 200
 201         list_for_each_entry_safe(rp, tmp, &lru_head, c_lru) {
 202                 if (!nfsd_cache_entry_expired(rp) &&
 203                     num_drc_entries <= max_drc_entries)
 204                         break;
 205                 nfsd_reply_cache_free_locked(rp);
 206         }
 207
 208         /*
 209          * Conditionally rearm the job. If we cleaned out the list, then
 210          * cancel any pending run (since there won't be any work to do).
 211          * Otherwise, we rearm the job or modify the existing one to run in
 212          * RC_EXPIRE since we just ran the pruner.
 213          */
 214         if (list_empty(&lru_head))
 215                 cancel_delayed_work(&cache_cleaner);
 216         else
 217                 mod_delayed_work(system_wq, &cache_cleaner, RC_EXPIRE);
 218 }
 219
 220 static void
 221 cache_cleaner_func(struct work_struct *unused)
 222 {
 223         spin_lock(&cache_lock);
 224         prune_cache_entries();
 225         spin_unlock(&cache_lock);
 226 }
 227
 228 static int
 229 nfsd_reply_cache_shrink(struct shrinker *shrink, struct shrink_control *sc)
 230 {
 231         unsigned int num;
 232
 233         spin_lock(&cache_lock);
 234         if (sc->nr_to_scan)
 235                 prune_cache_entries();
 236         num = num_drc_entries;
 237         spin_unlock(&cache_lock);
 238
 239         return num;
 240 }
 241
 242 /*
 243  * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
 244  */
 245 static __wsum
 246 nfsd_cache_csum(struct svc_rqst *rqstp)
 247 {
 248         int idx;
 249         unsigned int base;
 250         __wsum csum;
 251         struct xdr_buf *buf = &rqstp->rq_arg;
 252         const unsigned char *p = buf->head[0].iov_base;
 253         size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
 254                                 RC_CSUMLEN);
 255         size_t len = min(buf->head[0].iov_len, csum_len);
 256
 257         /* rq_arg.head first */
 258         csum = csum_partial(p, len, 0);
 259         csum_len -= len;
 260
 261         /* Continue into page array */
 262         idx = buf->page_base / PAGE_SIZE;
 263         base = buf->page_base & ~PAGE_MASK;
 264         while (csum_len) {
 265                 p = page_address(buf->pages[idx]) + base;
 266                 len = min_t(size_t, PAGE_SIZE - base, csum_len);
 267                 csum = csum_partial(p, len, csum);
 268                 csum_len -= len;
 269                 base = 0;
 270                 ++idx;
 271         }
 272         return csum;
 273 }
 274
 275 /*
 276  * Search the request hash for an entry that matches the given rqstp.
 277  * Must be called with cache_lock held. Returns the found entry or
 278  * NULL on failure.
 279  */
 280 static struct svc_cacherep *
 281 nfsd_cache_search(struct svc_rqst *rqstp, __wsum csum)
 282 {
 283         struct svc_cacherep     *rp;
 284         struct hlist_head       *rh;
 285         __be32                  xid = rqstp->rq_xid;
 286         u32                     proto =  rqstp->rq_prot,
 287                                 vers = rqstp->rq_vers,
 288                                 proc = rqstp->rq_proc;
 289
 290         rh = &cache_hash[request_hash(xid)];
 291         hlist_for_each_entry(rp, rh, c_hash) {
 292                 if (xid == rp->c_xid && proc == rp->c_proc &&
 293                     proto == rp->c_prot && vers == rp->c_vers &&
 294                     rqstp->rq_arg.len == rp->c_len && csum == rp->c_csum &&
 295                     rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) &&
 296                     rpc_get_port(svc_addr(rqstp)) == rpc_get_port((struct sockaddr *)&rp->c_addr))
 297                         return rp;
 298         }
 299         return NULL;
 300 }
 301
 302 /*
 303  * Try to find an entry matching the current call in the cache. When none
 304  * is found, we try to grab the oldest expired entry off the LRU list. If
 305  * a suitable one isn't there, then drop the cache_lock and allocate a
 306  * new one, then search again in case one got inserted while this thread
 307  * didn't hold the lock.
 308  */
 309 int
 310 nfsd_cache_lookup(struct svc_rqst *rqstp)
 311 {
 312         struct svc_cacherep     *rp, *found;
 313         __be32                  xid = rqstp->rq_xid;
 314         u32                     proto =  rqstp->rq_prot,
 315                                 vers = rqstp->rq_vers,
 316                                 proc = rqstp->rq_proc;
 317         __wsum                  csum;
 318         unsigned long           age;
 319         int type = rqstp->rq_cachetype;
 320         int rtn;
 321
 322         rqstp->rq_cacherep = NULL;
 323         if (type == RC_NOCACHE) {
 324                 nfsdstats.rcnocache++;
 325                 return RC_DOIT;
 326         }
 327
 328         csum = nfsd_cache_csum(rqstp);
 329
 330         spin_lock(&cache_lock);
 331         rtn = RC_DOIT;
 332
 333         rp = nfsd_cache_search(rqstp, csum);
 334         if (rp)
 335                 goto found_entry;
 336
 337         /* Try to use the first entry on the LRU */
 338         if (!list_empty(&lru_head)) {
 339                 rp = list_first_entry(&lru_head, struct svc_cacherep, c_lru);
 340                 if (nfsd_cache_entry_expired(rp) ||
 341                     num_drc_entries >= max_drc_entries) {
 342                         lru_put_end(rp);
 343                         prune_cache_entries();
 344                         goto setup_entry;
 345                 }
 346         }
 347
 348         /* Drop the lock and allocate a new entry */
 349         spin_unlock(&cache_lock);
 350         rp = nfsd_reply_cache_alloc();
 351         if (!rp) {
 352                 dprintk("nfsd: unable to allocate DRC entry!\n");
 353                 return RC_DOIT;
 354         }
 355         spin_lock(&cache_lock);
 356         ++num_drc_entries;
 357
 358         /*
 359          * Must search again just in case someone inserted one
 360          * after we dropped the lock above.
 361          */
 362         found = nfsd_cache_search(rqstp, csum);
 363         if (found) {
 364                 nfsd_reply_cache_free_locked(rp);
 365                 rp = found;
 366                 goto found_entry;
 367         }
 368
 369         /*
 370          * We're keeping the one we just allocated. Are we now over the
 371          * limit? Prune one off the tip of the LRU in trade for the one we
 372          * just allocated if so.
 373          */
 374         if (num_drc_entries >= max_drc_entries)
 375                 nfsd_reply_cache_free_locked(list_first_entry(&lru_head,
 376                                                 struct svc_cacherep, c_lru));
 377
 378 setup_entry:
 379         nfsdstats.rcmisses++;
 380         rqstp->rq_cacherep = rp;
 381         rp->c_state = RC_INPROG;
 382         rp->c_xid = xid;
 383         rp->c_proc = proc;
 384         rpc_copy_addr((struct sockaddr *)&rp->c_addr, svc_addr(rqstp));
 385         rpc_set_port((struct sockaddr *)&rp->c_addr, rpc_get_port(svc_addr(rqstp)));
 386         rp->c_prot = proto;
 387         rp->c_vers = vers;
 388         rp->c_len = rqstp->rq_arg.len;
 389         rp->c_csum = csum;
 390
 391         hash_refile(rp);
 392         lru_put_end(rp);
 393
 394         /* release any buffer */
 395         if (rp->c_type == RC_REPLBUFF) {
 396                 kfree(rp->c_replvec.iov_base);
 397                 rp->c_replvec.iov_base = NULL;
 398         }
 399         rp->c_type = RC_NOCACHE;
 400  out:
 401         spin_unlock(&cache_lock);
 402         return rtn;
 403
 404 found_entry:
 405         nfsdstats.rchits++;
 406         /* We found a matching entry which is either in progress or done. */
 407         age = jiffies - rp->c_timestamp;
 408         lru_put_end(rp);
 409
 410         rtn = RC_DROPIT;
 411         /* Request being processed or excessive rexmits */
 412         if (rp->c_state == RC_INPROG || age < RC_DELAY)
 413                 goto out;
 414
 415         /* From the hall of fame of impractical attacks:
 416          * Is this a user who tries to snoop on the cache? */
 417         rtn = RC_DOIT;
 418         if (!rqstp->rq_secure && rp->c_secure)
 419                 goto out;
 420
 421         /* Compose RPC reply header */
 422         switch (rp->c_type) {
 423         case RC_NOCACHE:
 424                 break;
 425         case RC_REPLSTAT:
 426                 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
 427                 rtn = RC_REPLY;
 428                 break;
 429         case RC_REPLBUFF:
 430                 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
 431                         goto out;       /* should not happen */
 432                 rtn = RC_REPLY;
 433                 break;
 434         default:
 435                 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
 436                 nfsd_reply_cache_free_locked(rp);
 437         }
 438
 439         goto out;
 440 }
 441
 442 /*
 443  * Update a cache entry. This is called from nfsd_dispatch when
 444  * the procedure has been executed and the complete reply is in
 445  * rqstp->rq_res.
 446  *
 447  * We're copying around data here rather than swapping buffers because
 448  * the toplevel loop requires max-sized buffers, which would be a waste
 449  * of memory for a cache with a max reply size of 100 bytes (diropokres).
 450  *
 451  * If we should start to use different types of cache entries tailored
 452  * specifically for attrstat and fh's, we may save even more space.
 453  *
 454  * Also note that a cachetype of RC_NOCACHE can legally be passed when
 455  * nfsd failed to encode a reply that otherwise would have been cached.
 456  * In this case, nfsd_cache_update is called with statp == NULL.
 457  */
 458 void
 459 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
 460 {
 461         struct svc_cacherep *rp = rqstp->rq_cacherep;
 462         struct kvec     *resv = &rqstp->rq_res.head[0], *cachv;
 463         int             len;
 464
 465         if (!rp)
 466                 return;
 467
 468         len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
 469         len >>= 2;
 470
 471         /* Don't cache excessive amounts of data and XDR failures */
 472         if (!statp || len > (256 >> 2)) {
 473                 nfsd_reply_cache_free(rp);
 474                 return;
 475         }
 476
 477         switch (cachetype) {
 478         case RC_REPLSTAT:
 479                 if (len != 1)
 480                         printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
 481                 rp->c_replstat = *statp;
 482                 break;
 483         case RC_REPLBUFF:
 484                 cachv = &rp->c_replvec;
 485                 cachv->iov_base = kmalloc(len << 2, GFP_KERNEL);
 486                 if (!cachv->iov_base) {
 487                         nfsd_reply_cache_free(rp);
 488                         return;
 489                 }
 490                 cachv->iov_len = len << 2;
 491                 memcpy(cachv->iov_base, statp, len << 2);
 492                 break;
 493         case RC_NOCACHE:
 494                 nfsd_reply_cache_free(rp);
 495                 return;
 496         }
 497         spin_lock(&cache_lock);
 498         lru_put_end(rp);
 499         rp->c_secure = rqstp->rq_secure;
 500         rp->c_type = cachetype;
 501         rp->c_state = RC_DONE;
 502         spin_unlock(&cache_lock);
 503         return;
 504 }
 505
 506 /*
 507  * Copy cached reply to current reply buffer. Should always fit.
 508  * FIXME as reply is in a page, we should just attach the page, and
 509  * keep a refcount....
 510  */
 511 static int
 512 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
 513 {
 514         struct kvec     *vec = &rqstp->rq_res.head[0];
 515
 516         if (vec->iov_len + data->iov_len > PAGE_SIZE) {
 517                 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
 518                                 data->iov_len);
 519                 return 0;
 520         }
 521         memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
 522         vec->iov_len += data->iov_len;
 523         return 1;
 524 }