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 <linux/log2.h>
  15 #include <linux/hash.h>
  16 #include <net/checksum.h>
  17
  18 #include "nfsd.h"
  19 #include "cache.h"
  20
  21 #define NFSDDBG_FACILITY        NFSDDBG_REPCACHE
  22
  23 /*
  24  * We use this value to determine the number of hash buckets from the max
  25  * cache size, the idea being that when the cache is at its maximum number
  26  * of entries, then this should be the average number of entries per bucket.
  27  */
  28 #define TARGET_BUCKET_SIZE      64
  29
  30 static struct hlist_head *      cache_hash;
  31 static struct list_head         lru_head;
  32 static struct kmem_cache        *drc_slab;
  33
  34 /* max number of entries allowed in the cache */
  35 static unsigned int             max_drc_entries;
  36
  37 /* number of significant bits in the hash value */
  38 static unsigned int             maskbits;
  39
  40 /*
  41  * Stats and other tracking of on the duplicate reply cache. All of these and
  42  * the "rc" fields in nfsdstats are protected by the cache_lock
  43  */
  44
  45 /* total number of entries */
  46 static unsigned int             num_drc_entries;
  47
  48 /* cache misses due only to checksum comparison failures */
  49 static unsigned int             payload_misses;
  50
  51 /* amount of memory (in bytes) currently consumed by the DRC */
  52 static unsigned int             drc_mem_usage;
  53
  54 /* longest hash chain seen */
  55 static unsigned int             longest_chain;
  56
  57 /* size of cache when we saw the longest hash chain */
  58 static unsigned int             longest_chain_cachesize;
  59
  60 static int      nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
  61 static void     cache_cleaner_func(struct work_struct *unused);
  62 static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
  63                                             struct shrink_control *sc);
  64 static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
  65                                            struct shrink_control *sc);
  66
  67 static struct shrinker nfsd_reply_cache_shrinker = {
  68         .scan_objects = nfsd_reply_cache_scan,
  69         .count_objects = nfsd_reply_cache_count,
  70         .seeks  = 1,
  71 };
  72
  73 /*
  74  * locking for the reply cache:
  75  * A cache entry is "single use" if c_state == RC_INPROG
  76  * Otherwise, it when accessing _prev or _next, the lock must be held.
  77  */
  78 static DEFINE_SPINLOCK(cache_lock);
  79 static DECLARE_DELAYED_WORK(cache_cleaner, cache_cleaner_func);
  80
  81 /*
  82  * Put a cap on the size of the DRC based on the amount of available
  83  * low memory in the machine.
  84  *
  85  *  64MB:    8192
  86  * 128MB:   11585
  87  * 256MB:   16384
  88  * 512MB:   23170
  89  *   1GB:   32768
  90  *   2GB:   46340
  91  *   4GB:   65536
  92  *   8GB:   92681
  93  *  16GB:  131072
  94  *
  95  * ...with a hard cap of 256k entries. In the worst case, each entry will be
  96  * ~1k, so the above numbers should give a rough max of the amount of memory
  97  * used in k.
  98  */
  99 static unsigned int
 100 nfsd_cache_size_limit(void)
 101 {
 102         unsigned int limit;
 103         unsigned long low_pages = totalram_pages - totalhigh_pages;
 104
 105         limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
 106         return min_t(unsigned int, limit, 256*1024);
 107 }
 108
 109 /*
 110  * Compute the number of hash buckets we need. Divide the max cachesize by
 111  * the "target" max bucket size, and round up to next power of two.
 112  */
 113 static unsigned int
 114 nfsd_hashsize(unsigned int limit)
 115 {
 116         return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
 117 }
 118
 119 static struct svc_cacherep *
 120 nfsd_reply_cache_alloc(void)
 121 {
 122         struct svc_cacherep     *rp;
 123
 124         rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
 125         if (rp) {
 126                 rp->c_state = RC_UNUSED;
 127                 rp->c_type = RC_NOCACHE;
 128                 INIT_LIST_HEAD(&rp->c_lru);
 129                 INIT_HLIST_NODE(&rp->c_hash);
 130         }
 131         return rp;
 132 }
 133
 134 static void
 135 nfsd_reply_cache_free_locked(struct svc_cacherep *rp)
 136 {
 137         if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
 138                 drc_mem_usage -= rp->c_replvec.iov_len;
 139                 kfree(rp->c_replvec.iov_base);
 140         }
 141         if (!hlist_unhashed(&rp->c_hash))
 142                 hlist_del(&rp->c_hash);
 143         list_del(&rp->c_lru);
 144         --num_drc_entries;
 145         drc_mem_usage -= sizeof(*rp);
 146         kmem_cache_free(drc_slab, rp);
 147 }
 148
 149 static void
 150 nfsd_reply_cache_free(struct svc_cacherep *rp)
 151 {
 152         spin_lock(&cache_lock);
 153         nfsd_reply_cache_free_locked(rp);
 154         spin_unlock(&cache_lock);
 155 }
 156
 157 int nfsd_reply_cache_init(void)
 158 {
 159         unsigned int hashsize;
 160
 161         INIT_LIST_HEAD(&lru_head);
 162         max_drc_entries = nfsd_cache_size_limit();
 163         num_drc_entries = 0;
 164         hashsize = nfsd_hashsize(max_drc_entries);
 165         maskbits = ilog2(hashsize);
 166
 167         register_shrinker(&nfsd_reply_cache_shrinker);
 168         drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
 169                                         0, 0, NULL);
 170         if (!drc_slab)
 171                 goto out_nomem;
 172
 173         cache_hash = kcalloc(hashsize, sizeof(struct hlist_head), GFP_KERNEL);
 174         if (!cache_hash)
 175                 goto out_nomem;
 176
 177         return 0;
 178 out_nomem:
 179         printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
 180         nfsd_reply_cache_shutdown();
 181         return -ENOMEM;
 182 }
 183
 184 void nfsd_reply_cache_shutdown(void)
 185 {
 186         struct svc_cacherep     *rp;
 187
 188         unregister_shrinker(&nfsd_reply_cache_shrinker);
 189         cancel_delayed_work_sync(&cache_cleaner);
 190
 191         while (!list_empty(&lru_head)) {
 192                 rp = list_entry(lru_head.next, struct svc_cacherep, c_lru);
 193                 nfsd_reply_cache_free_locked(rp);
 194         }
 195
 196         kfree (cache_hash);
 197         cache_hash = NULL;
 198
 199         if (drc_slab) {
 200                 kmem_cache_destroy(drc_slab);
 201                 drc_slab = NULL;
 202         }
 203 }
 204
 205 /*
 206  * Move cache entry to end of LRU list, and queue the cleaner to run if it's
 207  * not already scheduled.
 208  */
 209 static void
 210 lru_put_end(struct svc_cacherep *rp)
 211 {
 212         rp->c_timestamp = jiffies;
 213         list_move_tail(&rp->c_lru, &lru_head);
 214         schedule_delayed_work(&cache_cleaner, RC_EXPIRE);
 215 }
 216
 217 /*
 218  * Move a cache entry from one hash list to another
 219  */
 220 static void
 221 hash_refile(struct svc_cacherep *rp)
 222 {
 223         hlist_del_init(&rp->c_hash);
 224         hlist_add_head(&rp->c_hash, cache_hash + hash_32(rp->c_xid, maskbits));
 225 }
 226
 227 static inline bool
 228 nfsd_cache_entry_expired(struct svc_cacherep *rp)
 229 {
 230         return rp->c_state != RC_INPROG &&
 231                time_after(jiffies, rp->c_timestamp + RC_EXPIRE);
 232 }
 233
 234 /*
 235  * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
 236  * Also prune the oldest ones when the total exceeds the max number of entries.
 237  */
 238 static long
 239 prune_cache_entries(void)
 240 {
 241         struct svc_cacherep *rp, *tmp;
 242         long freed = 0;
 243
 244         list_for_each_entry_safe(rp, tmp, &lru_head, c_lru) {
 245                 if (!nfsd_cache_entry_expired(rp) &&
 246                     num_drc_entries <= max_drc_entries)
 247                         break;
 248                 nfsd_reply_cache_free_locked(rp);
 249                 freed++;
 250         }
 251
 252         /*
 253          * Conditionally rearm the job. If we cleaned out the list, then
 254          * cancel any pending run (since there won't be any work to do).
 255          * Otherwise, we rearm the job or modify the existing one to run in
 256          * RC_EXPIRE since we just ran the pruner.
 257          */
 258         if (list_empty(&lru_head))
 259                 cancel_delayed_work(&cache_cleaner);
 260         else
 261                 mod_delayed_work(system_wq, &cache_cleaner, RC_EXPIRE);
 262         return freed;
 263 }
 264
 265 static void
 266 cache_cleaner_func(struct work_struct *unused)
 267 {
 268         spin_lock(&cache_lock);
 269         prune_cache_entries();
 270         spin_unlock(&cache_lock);
 271 }
 272
 273 static unsigned long
 274 nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
 275 {
 276         unsigned long num;
 277
 278         spin_lock(&cache_lock);
 279         num = num_drc_entries;
 280         spin_unlock(&cache_lock);
 281
 282         return num;
 283 }
 284
 285 static unsigned long
 286 nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
 287 {
 288         unsigned long freed;
 289
 290         spin_lock(&cache_lock);
 291         freed = prune_cache_entries();
 292         spin_unlock(&cache_lock);
 293         return freed;
 294 }
 295 /*
 296  * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
 297  */
 298 static __wsum
 299 nfsd_cache_csum(struct svc_rqst *rqstp)
 300 {
 301         int idx;
 302         unsigned int base;
 303         __wsum csum;
 304         struct xdr_buf *buf = &rqstp->rq_arg;
 305         const unsigned char *p = buf->head[0].iov_base;
 306         size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
 307                                 RC_CSUMLEN);
 308         size_t len = min(buf->head[0].iov_len, csum_len);
 309
 310         /* rq_arg.head first */
 311         csum = csum_partial(p, len, 0);
 312         csum_len -= len;
 313
 314         /* Continue into page array */
 315         idx = buf->page_base / PAGE_SIZE;
 316         base = buf->page_base & ~PAGE_MASK;
 317         while (csum_len) {
 318                 p = page_address(buf->pages[idx]) + base;
 319                 len = min_t(size_t, PAGE_SIZE - base, csum_len);
 320                 csum = csum_partial(p, len, csum);
 321                 csum_len -= len;
 322                 base = 0;
 323                 ++idx;
 324         }
 325         return csum;
 326 }
 327
 328 static bool
 329 nfsd_cache_match(struct svc_rqst *rqstp, __wsum csum, struct svc_cacherep *rp)
 330 {
 331         /* Check RPC header info first */
 332         if (rqstp->rq_xid != rp->c_xid || rqstp->rq_proc != rp->c_proc ||
 333             rqstp->rq_prot != rp->c_prot || rqstp->rq_vers != rp->c_vers ||
 334             rqstp->rq_arg.len != rp->c_len ||
 335             !rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) ||
 336             rpc_get_port(svc_addr(rqstp)) != rpc_get_port((struct sockaddr *)&rp->c_addr))
 337                 return false;
 338
 339         /* compare checksum of NFS data */
 340         if (csum != rp->c_csum) {
 341                 ++payload_misses;
 342                 return false;
 343         }
 344
 345         return true;
 346 }
 347
 348 /*
 349  * Search the request hash for an entry that matches the given rqstp.
 350  * Must be called with cache_lock held. Returns the found entry or
 351  * NULL on failure.
 352  */
 353 static struct svc_cacherep *
 354 nfsd_cache_search(struct svc_rqst *rqstp, __wsum csum)
 355 {
 356         struct svc_cacherep     *rp, *ret = NULL;
 357         struct hlist_head       *rh;
 358         unsigned int            entries = 0;
 359
 360         rh = &cache_hash[hash_32(rqstp->rq_xid, maskbits)];
 361         hlist_for_each_entry(rp, rh, c_hash) {
 362                 ++entries;
 363                 if (nfsd_cache_match(rqstp, csum, rp)) {
 364                         ret = rp;
 365                         break;
 366                 }
 367         }
 368
 369         /* tally hash chain length stats */
 370         if (entries > longest_chain) {
 371                 longest_chain = entries;
 372                 longest_chain_cachesize = num_drc_entries;
 373         } else if (entries == longest_chain) {
 374                 /* prefer to keep the smallest cachesize possible here */
 375                 longest_chain_cachesize = min(longest_chain_cachesize,
 376                                                 num_drc_entries);
 377         }
 378
 379         return ret;
 380 }
 381
 382 /*
 383  * Try to find an entry matching the current call in the cache. When none
 384  * is found, we try to grab the oldest expired entry off the LRU list. If
 385  * a suitable one isn't there, then drop the cache_lock and allocate a
 386  * new one, then search again in case one got inserted while this thread
 387  * didn't hold the lock.
 388  */
 389 int
 390 nfsd_cache_lookup(struct svc_rqst *rqstp)
 391 {
 392         struct svc_cacherep     *rp, *found;
 393         __be32                  xid = rqstp->rq_xid;
 394         u32                     proto =  rqstp->rq_prot,
 395                                 vers = rqstp->rq_vers,
 396                                 proc = rqstp->rq_proc;
 397         __wsum                  csum;
 398         unsigned long           age;
 399         int type = rqstp->rq_cachetype;
 400         int rtn = RC_DOIT;
 401
 402         rqstp->rq_cacherep = NULL;
 403         if (type == RC_NOCACHE) {
 404                 nfsdstats.rcnocache++;
 405                 return rtn;
 406         }
 407
 408         csum = nfsd_cache_csum(rqstp);
 409
 410         /*
 411          * Since the common case is a cache miss followed by an insert,
 412          * preallocate an entry. First, try to reuse the first entry on the LRU
 413          * if it works, then go ahead and prune the LRU list.
 414          */
 415         spin_lock(&cache_lock);
 416         if (!list_empty(&lru_head)) {
 417                 rp = list_first_entry(&lru_head, struct svc_cacherep, c_lru);
 418                 if (nfsd_cache_entry_expired(rp) ||
 419                     num_drc_entries >= max_drc_entries) {
 420                         lru_put_end(rp);
 421                         prune_cache_entries();
 422                         goto search_cache;
 423                 }
 424         }
 425
 426         /* No expired ones available, allocate a new one. */
 427         spin_unlock(&cache_lock);
 428         rp = nfsd_reply_cache_alloc();
 429         spin_lock(&cache_lock);
 430         if (likely(rp)) {
 431                 ++num_drc_entries;
 432                 drc_mem_usage += sizeof(*rp);
 433         }
 434
 435 search_cache:
 436         found = nfsd_cache_search(rqstp, csum);
 437         if (found) {
 438                 if (likely(rp))
 439                         nfsd_reply_cache_free_locked(rp);
 440                 rp = found;
 441                 goto found_entry;
 442         }
 443
 444         if (!rp) {
 445                 dprintk("nfsd: unable to allocate DRC entry!\n");
 446                 goto out;
 447         }
 448
 449         /*
 450          * We're keeping the one we just allocated. Are we now over the
 451          * limit? Prune one off the tip of the LRU in trade for the one we
 452          * just allocated if so.
 453          */
 454         if (num_drc_entries >= max_drc_entries)
 455                 nfsd_reply_cache_free_locked(list_first_entry(&lru_head,
 456                                                 struct svc_cacherep, c_lru));
 457
 458         nfsdstats.rcmisses++;
 459         rqstp->rq_cacherep = rp;
 460         rp->c_state = RC_INPROG;
 461         rp->c_xid = xid;
 462         rp->c_proc = proc;
 463         rpc_copy_addr((struct sockaddr *)&rp->c_addr, svc_addr(rqstp));
 464         rpc_set_port((struct sockaddr *)&rp->c_addr, rpc_get_port(svc_addr(rqstp)));
 465         rp->c_prot = proto;
 466         rp->c_vers = vers;
 467         rp->c_len = rqstp->rq_arg.len;
 468         rp->c_csum = csum;
 469
 470         hash_refile(rp);
 471         lru_put_end(rp);
 472
 473         /* release any buffer */
 474         if (rp->c_type == RC_REPLBUFF) {
 475                 drc_mem_usage -= rp->c_replvec.iov_len;
 476                 kfree(rp->c_replvec.iov_base);
 477                 rp->c_replvec.iov_base = NULL;
 478         }
 479         rp->c_type = RC_NOCACHE;
 480  out:
 481         spin_unlock(&cache_lock);
 482         return rtn;
 483
 484 found_entry:
 485         nfsdstats.rchits++;
 486         /* We found a matching entry which is either in progress or done. */
 487         age = jiffies - rp->c_timestamp;
 488         lru_put_end(rp);
 489
 490         rtn = RC_DROPIT;
 491         /* Request being processed or excessive rexmits */
 492         if (rp->c_state == RC_INPROG || age < RC_DELAY)
 493                 goto out;
 494
 495         /* From the hall of fame of impractical attacks:
 496          * Is this a user who tries to snoop on the cache? */
 497         rtn = RC_DOIT;
 498         if (!rqstp->rq_secure && rp->c_secure)
 499                 goto out;
 500
 501         /* Compose RPC reply header */
 502         switch (rp->c_type) {
 503         case RC_NOCACHE:
 504                 break;
 505         case RC_REPLSTAT:
 506                 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
 507                 rtn = RC_REPLY;
 508                 break;
 509         case RC_REPLBUFF:
 510                 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
 511                         goto out;       /* should not happen */
 512                 rtn = RC_REPLY;
 513                 break;
 514         default:
 515                 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
 516                 nfsd_reply_cache_free_locked(rp);
 517         }
 518
 519         goto out;
 520 }
 521
 522 /*
 523  * Update a cache entry. This is called from nfsd_dispatch when
 524  * the procedure has been executed and the complete reply is in
 525  * rqstp->rq_res.
 526  *
 527  * We're copying around data here rather than swapping buffers because
 528  * the toplevel loop requires max-sized buffers, which would be a waste
 529  * of memory for a cache with a max reply size of 100 bytes (diropokres).
 530  *
 531  * If we should start to use different types of cache entries tailored
 532  * specifically for attrstat and fh's, we may save even more space.
 533  *
 534  * Also note that a cachetype of RC_NOCACHE can legally be passed when
 535  * nfsd failed to encode a reply that otherwise would have been cached.
 536  * In this case, nfsd_cache_update is called with statp == NULL.
 537  */
 538 void
 539 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
 540 {
 541         struct svc_cacherep *rp = rqstp->rq_cacherep;
 542         struct kvec     *resv = &rqstp->rq_res.head[0], *cachv;
 543         int             len;
 544         size_t          bufsize = 0;
 545
 546         if (!rp)
 547                 return;
 548
 549         len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
 550         len >>= 2;
 551
 552         /* Don't cache excessive amounts of data and XDR failures */
 553         if (!statp || len > (256 >> 2)) {
 554                 nfsd_reply_cache_free(rp);
 555                 return;
 556         }
 557
 558         switch (cachetype) {
 559         case RC_REPLSTAT:
 560                 if (len != 1)
 561                         printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
 562                 rp->c_replstat = *statp;
 563                 break;
 564         case RC_REPLBUFF:
 565                 cachv = &rp->c_replvec;
 566                 bufsize = len << 2;
 567                 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
 568                 if (!cachv->iov_base) {
 569                         nfsd_reply_cache_free(rp);
 570                         return;
 571                 }
 572                 cachv->iov_len = bufsize;
 573                 memcpy(cachv->iov_base, statp, bufsize);
 574                 break;
 575         case RC_NOCACHE:
 576                 nfsd_reply_cache_free(rp);
 577                 return;
 578         }
 579         spin_lock(&cache_lock);
 580         drc_mem_usage += bufsize;
 581         lru_put_end(rp);
 582         rp->c_secure = rqstp->rq_secure;
 583         rp->c_type = cachetype;
 584         rp->c_state = RC_DONE;
 585         spin_unlock(&cache_lock);
 586         return;
 587 }
 588
 589 /*
 590  * Copy cached reply to current reply buffer. Should always fit.
 591  * FIXME as reply is in a page, we should just attach the page, and
 592  * keep a refcount....
 593  */
 594 static int
 595 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
 596 {
 597         struct kvec     *vec = &rqstp->rq_res.head[0];
 598
 599         if (vec->iov_len + data->iov_len > PAGE_SIZE) {
 600                 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
 601                                 data->iov_len);
 602                 return 0;
 603         }
 604         memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
 605         vec->iov_len += data->iov_len;
 606         return 1;
 607 }
 608
 609 /*
 610  * Note that fields may be added, removed or reordered in the future. Programs
 611  * scraping this file for info should test the labels to ensure they're
 612  * getting the correct field.
 613  */
 614 static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
 615 {
 616         spin_lock(&cache_lock);
 617         seq_printf(m, "max entries:           %u\n", max_drc_entries);
 618         seq_printf(m, "num entries:           %u\n", num_drc_entries);
 619         seq_printf(m, "hash buckets:          %u\n", 1 << maskbits);
 620         seq_printf(m, "mem usage:             %u\n", drc_mem_usage);
 621         seq_printf(m, "cache hits:            %u\n", nfsdstats.rchits);
 622         seq_printf(m, "cache misses:          %u\n", nfsdstats.rcmisses);
 623         seq_printf(m, "not cached:            %u\n", nfsdstats.rcnocache);
 624         seq_printf(m, "payload misses:        %u\n", payload_misses);
 625         seq_printf(m, "longest chain len:     %u\n", longest_chain);
 626         seq_printf(m, "cachesize at longest:  %u\n", longest_chain_cachesize);
 627         spin_unlock(&cache_lock);
 628         return 0;
 629 }
 630
 631 int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
 632 {
 633         return single_open(file, nfsd_reply_cache_stats_show, NULL);
 634 }