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         if (!hlist_unhashed(&rp->c_hash))
 106                 hlist_del(&rp->c_hash);
 107         list_del(&rp->c_lru);
 108         --num_drc_entries;
 109         kmem_cache_free(drc_slab, rp);
 110 }
 111
 112 static void
 113 nfsd_reply_cache_free(struct svc_cacherep *rp)
 114 {
 115         spin_lock(&cache_lock);
 116         nfsd_reply_cache_free_locked(rp);
 117         spin_unlock(&cache_lock);
 118 }
 119
 120 int nfsd_reply_cache_init(void)
 121 {
 122         INIT_LIST_HEAD(&lru_head);
 123         max_drc_entries = nfsd_cache_size_limit();
 124         num_drc_entries = 0;
 125
 126         register_shrinker(&nfsd_reply_cache_shrinker);
 127         drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
 128                                         0, 0, NULL);
 129         if (!drc_slab)
 130                 goto out_nomem;
 131
 132         cache_hash = kcalloc(HASHSIZE, sizeof(struct hlist_head), GFP_KERNEL);
 133         if (!cache_hash)
 134                 goto out_nomem;
 135
 136         return 0;
 137 out_nomem:
 138         printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
 139         nfsd_reply_cache_shutdown();
 140         return -ENOMEM;
 141 }
 142
 143 void nfsd_reply_cache_shutdown(void)
 144 {
 145         struct svc_cacherep     *rp;
 146
 147         unregister_shrinker(&nfsd_reply_cache_shrinker);
 148         cancel_delayed_work_sync(&cache_cleaner);
 149
 150         while (!list_empty(&lru_head)) {
 151                 rp = list_entry(lru_head.next, struct svc_cacherep, c_lru);
 152                 nfsd_reply_cache_free_locked(rp);
 153         }
 154
 155         kfree (cache_hash);
 156         cache_hash = NULL;
 157
 158         if (drc_slab) {
 159                 kmem_cache_destroy(drc_slab);
 160                 drc_slab = NULL;
 161         }
 162 }
 163
 164 /*
 165  * Move cache entry to end of LRU list, and queue the cleaner to run if it's
 166  * not already scheduled.
 167  */
 168 static void
 169 lru_put_end(struct svc_cacherep *rp)
 170 {
 171         rp->c_timestamp = jiffies;
 172         list_move_tail(&rp->c_lru, &lru_head);
 173         schedule_delayed_work(&cache_cleaner, RC_EXPIRE);
 174 }
 175
 176 /*
 177  * Move a cache entry from one hash list to another
 178  */
 179 static void
 180 hash_refile(struct svc_cacherep *rp)
 181 {
 182         hlist_del_init(&rp->c_hash);
 183         hlist_add_head(&rp->c_hash, cache_hash + request_hash(rp->c_xid));
 184 }
 185
 186 static inline bool
 187 nfsd_cache_entry_expired(struct svc_cacherep *rp)
 188 {
 189         return rp->c_state != RC_INPROG &&
 190                time_after(jiffies, rp->c_timestamp + RC_EXPIRE);
 191 }
 192
 193 /*
 194  * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
 195  * Also prune the oldest ones when the total exceeds the max number of entries.
 196  */
 197 static void
 198 prune_cache_entries(void)
 199 {
 200         struct svc_cacherep *rp, *tmp;
 201
 202         list_for_each_entry_safe(rp, tmp, &lru_head, c_lru) {
 203                 if (!nfsd_cache_entry_expired(rp) &&
 204                     num_drc_entries <= max_drc_entries)
 205                         break;
 206                 nfsd_reply_cache_free_locked(rp);
 207         }
 208
 209         /*
 210          * Conditionally rearm the job. If we cleaned out the list, then
 211          * cancel any pending run (since there won't be any work to do).
 212          * Otherwise, we rearm the job or modify the existing one to run in
 213          * RC_EXPIRE since we just ran the pruner.
 214          */
 215         if (list_empty(&lru_head))
 216                 cancel_delayed_work(&cache_cleaner);
 217         else
 218                 mod_delayed_work(system_wq, &cache_cleaner, RC_EXPIRE);
 219 }
 220
 221 static void
 222 cache_cleaner_func(struct work_struct *unused)
 223 {
 224         spin_lock(&cache_lock);
 225         prune_cache_entries();
 226         spin_unlock(&cache_lock);
 227 }
 228
 229 static int
 230 nfsd_reply_cache_shrink(struct shrinker *shrink, struct shrink_control *sc)
 231 {
 232         unsigned int num;
 233
 234         spin_lock(&cache_lock);
 235         if (sc->nr_to_scan)
 236                 prune_cache_entries();
 237         num = num_drc_entries;
 238         spin_unlock(&cache_lock);
 239
 240         return num;
 241 }
 242
 243 /*
 244  * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
 245  */
 246 static __wsum
 247 nfsd_cache_csum(struct svc_rqst *rqstp)
 248 {
 249         int idx;
 250         unsigned int base;
 251         __wsum csum;
 252         struct xdr_buf *buf = &rqstp->rq_arg;
 253         const unsigned char *p = buf->head[0].iov_base;
 254         size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
 255                                 RC_CSUMLEN);
 256         size_t len = min(buf->head[0].iov_len, csum_len);
 257
 258         /* rq_arg.head first */
 259         csum = csum_partial(p, len, 0);
 260         csum_len -= len;
 261
 262         /* Continue into page array */
 263         idx = buf->page_base / PAGE_SIZE;
 264         base = buf->page_base & ~PAGE_MASK;
 265         while (csum_len) {
 266                 p = page_address(buf->pages[idx]) + base;
 267                 len = min_t(size_t, PAGE_SIZE - base, csum_len);
 268                 csum = csum_partial(p, len, csum);
 269                 csum_len -= len;
 270                 base = 0;
 271                 ++idx;
 272         }
 273         return csum;
 274 }
 275
 276 /*
 277  * Search the request hash for an entry that matches the given rqstp.
 278  * Must be called with cache_lock held. Returns the found entry or
 279  * NULL on failure.
 280  */
 281 static struct svc_cacherep *
 282 nfsd_cache_search(struct svc_rqst *rqstp, __wsum csum)
 283 {
 284         struct svc_cacherep     *rp;
 285         struct hlist_head       *rh;
 286         __be32                  xid = rqstp->rq_xid;
 287         u32                     proto =  rqstp->rq_prot,
 288                                 vers = rqstp->rq_vers,
 289                                 proc = rqstp->rq_proc;
 290
 291         rh = &cache_hash[request_hash(xid)];
 292         hlist_for_each_entry(rp, rh, c_hash) {
 293                 if (xid == rp->c_xid && proc == rp->c_proc &&
 294                     proto == rp->c_prot && vers == rp->c_vers &&
 295                     rqstp->rq_arg.len == rp->c_len && csum == rp->c_csum &&
 296                     rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) &&
 297                     rpc_get_port(svc_addr(rqstp)) == rpc_get_port((struct sockaddr *)&rp->c_addr))
 298                         return rp;
 299         }
 300         return NULL;
 301 }
 302
 303 /*
 304  * Try to find an entry matching the current call in the cache. When none
 305  * is found, we try to grab the oldest expired entry off the LRU list. If
 306  * a suitable one isn't there, then drop the cache_lock and allocate a
 307  * new one, then search again in case one got inserted while this thread
 308  * didn't hold the lock.
 309  */
 310 int
 311 nfsd_cache_lookup(struct svc_rqst *rqstp)
 312 {
 313         struct svc_cacherep     *rp, *found;
 314         __be32                  xid = rqstp->rq_xid;
 315         u32                     proto =  rqstp->rq_prot,
 316                                 vers = rqstp->rq_vers,
 317                                 proc = rqstp->rq_proc;
 318         __wsum                  csum;
 319         unsigned long           age;
 320         int type = rqstp->rq_cachetype;
 321         int rtn;
 322
 323         rqstp->rq_cacherep = NULL;
 324         if (type == RC_NOCACHE) {
 325                 nfsdstats.rcnocache++;
 326                 return RC_DOIT;
 327         }
 328
 329         csum = nfsd_cache_csum(rqstp);
 330
 331         spin_lock(&cache_lock);
 332         rtn = RC_DOIT;
 333
 334         rp = nfsd_cache_search(rqstp, csum);
 335         if (rp)
 336                 goto found_entry;
 337
 338         /* Try to use the first entry on the LRU */
 339         if (!list_empty(&lru_head)) {
 340                 rp = list_first_entry(&lru_head, struct svc_cacherep, c_lru);
 341                 if (nfsd_cache_entry_expired(rp) ||
 342                     num_drc_entries >= max_drc_entries) {
 343                         lru_put_end(rp);
 344                         prune_cache_entries();
 345                         goto setup_entry;
 346                 }
 347         }
 348
 349         /* Drop the lock and allocate a new entry */
 350         spin_unlock(&cache_lock);
 351         rp = nfsd_reply_cache_alloc();
 352         if (!rp) {
 353                 dprintk("nfsd: unable to allocate DRC entry!\n");
 354                 return RC_DOIT;
 355         }
 356         spin_lock(&cache_lock);
 357         ++num_drc_entries;
 358
 359         /*
 360          * Must search again just in case someone inserted one
 361          * after we dropped the lock above.
 362          */
 363         found = nfsd_cache_search(rqstp, csum);
 364         if (found) {
 365                 nfsd_reply_cache_free_locked(rp);
 366                 rp = found;
 367                 goto found_entry;
 368         }
 369
 370         /*
 371          * We're keeping the one we just allocated. Are we now over the
 372          * limit? Prune one off the tip of the LRU in trade for the one we
 373          * just allocated if so.
 374          */
 375         if (num_drc_entries >= max_drc_entries)
 376                 nfsd_reply_cache_free_locked(list_first_entry(&lru_head,
 377                                                 struct svc_cacherep, c_lru));
 378
 379 setup_entry:
 380         nfsdstats.rcmisses++;
 381         rqstp->rq_cacherep = rp;
 382         rp->c_state = RC_INPROG;
 383         rp->c_xid = xid;
 384         rp->c_proc = proc;
 385         rpc_copy_addr((struct sockaddr *)&rp->c_addr, svc_addr(rqstp));
 386         rpc_set_port((struct sockaddr *)&rp->c_addr, rpc_get_port(svc_addr(rqstp)));
 387         rp->c_prot = proto;
 388         rp->c_vers = vers;
 389         rp->c_len = rqstp->rq_arg.len;
 390         rp->c_csum = csum;
 391
 392         hash_refile(rp);
 393         lru_put_end(rp);
 394
 395         /* release any buffer */
 396         if (rp->c_type == RC_REPLBUFF) {
 397                 kfree(rp->c_replvec.iov_base);
 398                 rp->c_replvec.iov_base = NULL;
 399         }
 400         rp->c_type = RC_NOCACHE;
 401  out:
 402         spin_unlock(&cache_lock);
 403         return rtn;
 404
 405 found_entry:
 406         nfsdstats.rchits++;
 407         /* We found a matching entry which is either in progress or done. */
 408         age = jiffies - rp->c_timestamp;
 409         lru_put_end(rp);
 410
 411         rtn = RC_DROPIT;
 412         /* Request being processed or excessive rexmits */
 413         if (rp->c_state == RC_INPROG || age < RC_DELAY)
 414                 goto out;
 415
 416         /* From the hall of fame of impractical attacks:
 417          * Is this a user who tries to snoop on the cache? */
 418         rtn = RC_DOIT;
 419         if (!rqstp->rq_secure && rp->c_secure)
 420                 goto out;
 421
 422         /* Compose RPC reply header */
 423         switch (rp->c_type) {
 424         case RC_NOCACHE:
 425                 break;
 426         case RC_REPLSTAT:
 427                 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
 428                 rtn = RC_REPLY;
 429                 break;
 430         case RC_REPLBUFF:
 431                 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
 432                         goto out;       /* should not happen */
 433                 rtn = RC_REPLY;
 434                 break;
 435         default:
 436                 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
 437                 nfsd_reply_cache_free_locked(rp);
 438         }
 439
 440         goto out;
 441 }
 442
 443 /*
 444  * Update a cache entry. This is called from nfsd_dispatch when
 445  * the procedure has been executed and the complete reply is in
 446  * rqstp->rq_res.
 447  *
 448  * We're copying around data here rather than swapping buffers because
 449  * the toplevel loop requires max-sized buffers, which would be a waste
 450  * of memory for a cache with a max reply size of 100 bytes (diropokres).
 451  *
 452  * If we should start to use different types of cache entries tailored
 453  * specifically for attrstat and fh's, we may save even more space.
 454  *
 455  * Also note that a cachetype of RC_NOCACHE can legally be passed when
 456  * nfsd failed to encode a reply that otherwise would have been cached.
 457  * In this case, nfsd_cache_update is called with statp == NULL.
 458  */
 459 void
 460 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
 461 {
 462         struct svc_cacherep *rp = rqstp->rq_cacherep;
 463         struct kvec     *resv = &rqstp->rq_res.head[0], *cachv;
 464         int             len;
 465
 466         if (!rp)
 467                 return;
 468
 469         len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
 470         len >>= 2;
 471
 472         /* Don't cache excessive amounts of data and XDR failures */
 473         if (!statp || len > (256 >> 2)) {
 474                 nfsd_reply_cache_free(rp);
 475                 return;
 476         }
 477
 478         switch (cachetype) {
 479         case RC_REPLSTAT:
 480                 if (len != 1)
 481                         printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
 482                 rp->c_replstat = *statp;
 483                 break;
 484         case RC_REPLBUFF:
 485                 cachv = &rp->c_replvec;
 486                 cachv->iov_base = kmalloc(len << 2, GFP_KERNEL);
 487                 if (!cachv->iov_base) {
 488                         nfsd_reply_cache_free(rp);
 489                         return;
 490                 }
 491                 cachv->iov_len = len << 2;
 492                 memcpy(cachv->iov_base, statp, len << 2);
 493                 break;
 494         case RC_NOCACHE:
 495                 nfsd_reply_cache_free(rp);
 496                 return;
 497         }
 498         spin_lock(&cache_lock);
 499         lru_put_end(rp);
 500         rp->c_secure = rqstp->rq_secure;
 501         rp->c_type = cachetype;
 502         rp->c_state = RC_DONE;
 503         spin_unlock(&cache_lock);
 504         return;
 505 }
 506
 507 /*
 508  * Copy cached reply to current reply buffer. Should always fit.
 509  * FIXME as reply is in a page, we should just attach the page, and
 510  * keep a refcount....
 511  */
 512 static int
 513 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
 514 {
 515         struct kvec     *vec = &rqstp->rq_res.head[0];
 516
 517         if (vec->iov_len + data->iov_len > PAGE_SIZE) {
 518                 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
 519                                 data->iov_len);
 520                 return 0;
 521         }
 522         memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
 523         vec->iov_len += data->iov_len;
 524         return 1;
 525 }