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         /*
 225          * No point in byte swapping c_xid since we're just using it to pick
 226          * a hash bucket.
 227          */
 228         hlist_add_head(&rp->c_hash, cache_hash +
 229                         hash_32((__force u32)rp->c_xid, maskbits));
 230 }
 231
 232 /*
 233  * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
 234  * Also prune the oldest ones when the total exceeds the max number of entries.
 235  */
 236 static long
 237 prune_cache_entries(void)
 238 {
 239         struct svc_cacherep *rp, *tmp;
 240         long freed = 0;
 241
 242         list_for_each_entry_safe(rp, tmp, &lru_head, c_lru) {
 243                 /*
 244                  * Don't free entries attached to calls that are still
 245                  * in-progress, but do keep scanning the list.
 246                  */
 247                 if (rp->c_state == RC_INPROG)
 248                         continue;
 249                 if (num_drc_entries <= max_drc_entries &&
 250                     time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
 251                         break;
 252                 nfsd_reply_cache_free_locked(rp);
 253                 freed++;
 254         }
 255
 256         /*
 257          * Conditionally rearm the job. If we cleaned out the list, then
 258          * cancel any pending run (since there won't be any work to do).
 259          * Otherwise, we rearm the job or modify the existing one to run in
 260          * RC_EXPIRE since we just ran the pruner.
 261          */
 262         if (list_empty(&lru_head))
 263                 cancel_delayed_work(&cache_cleaner);
 264         else
 265                 mod_delayed_work(system_wq, &cache_cleaner, RC_EXPIRE);
 266         return freed;
 267 }
 268
 269 static void
 270 cache_cleaner_func(struct work_struct *unused)
 271 {
 272         spin_lock(&cache_lock);
 273         prune_cache_entries();
 274         spin_unlock(&cache_lock);
 275 }
 276
 277 static unsigned long
 278 nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
 279 {
 280         unsigned long num;
 281
 282         spin_lock(&cache_lock);
 283         num = num_drc_entries;
 284         spin_unlock(&cache_lock);
 285
 286         return num;
 287 }
 288
 289 static unsigned long
 290 nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
 291 {
 292         unsigned long freed;
 293
 294         spin_lock(&cache_lock);
 295         freed = prune_cache_entries();
 296         spin_unlock(&cache_lock);
 297         return freed;
 298 }
 299 /*
 300  * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
 301  */
 302 static __wsum
 303 nfsd_cache_csum(struct svc_rqst *rqstp)
 304 {
 305         int idx;
 306         unsigned int base;
 307         __wsum csum;
 308         struct xdr_buf *buf = &rqstp->rq_arg;
 309         const unsigned char *p = buf->head[0].iov_base;
 310         size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
 311                                 RC_CSUMLEN);
 312         size_t len = min(buf->head[0].iov_len, csum_len);
 313
 314         /* rq_arg.head first */
 315         csum = csum_partial(p, len, 0);
 316         csum_len -= len;
 317
 318         /* Continue into page array */
 319         idx = buf->page_base / PAGE_SIZE;
 320         base = buf->page_base & ~PAGE_MASK;
 321         while (csum_len) {
 322                 p = page_address(buf->pages[idx]) + base;
 323                 len = min_t(size_t, PAGE_SIZE - base, csum_len);
 324                 csum = csum_partial(p, len, csum);
 325                 csum_len -= len;
 326                 base = 0;
 327                 ++idx;
 328         }
 329         return csum;
 330 }
 331
 332 static bool
 333 nfsd_cache_match(struct svc_rqst *rqstp, __wsum csum, struct svc_cacherep *rp)
 334 {
 335         /* Check RPC header info first */
 336         if (rqstp->rq_xid != rp->c_xid || rqstp->rq_proc != rp->c_proc ||
 337             rqstp->rq_prot != rp->c_prot || rqstp->rq_vers != rp->c_vers ||
 338             rqstp->rq_arg.len != rp->c_len ||
 339             !rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) ||
 340             rpc_get_port(svc_addr(rqstp)) != rpc_get_port((struct sockaddr *)&rp->c_addr))
 341                 return false;
 342
 343         /* compare checksum of NFS data */
 344         if (csum != rp->c_csum) {
 345                 ++payload_misses;
 346                 return false;
 347         }
 348
 349         return true;
 350 }
 351
 352 /*
 353  * Search the request hash for an entry that matches the given rqstp.
 354  * Must be called with cache_lock held. Returns the found entry or
 355  * NULL on failure.
 356  */
 357 static struct svc_cacherep *
 358 nfsd_cache_search(struct svc_rqst *rqstp, __wsum csum)
 359 {
 360         struct svc_cacherep     *rp, *ret = NULL;
 361         struct hlist_head       *rh;
 362         unsigned int            entries = 0;
 363
 364         /*
 365          * No point in byte swapping rq_xid since we're just using it to pick
 366          * a hash bucket.
 367          */
 368         rh = &cache_hash[hash_32((__force u32)rqstp->rq_xid, maskbits)];
 369         hlist_for_each_entry(rp, rh, c_hash) {
 370                 ++entries;
 371                 if (nfsd_cache_match(rqstp, csum, rp)) {
 372                         ret = rp;
 373                         break;
 374                 }
 375         }
 376
 377         /* tally hash chain length stats */
 378         if (entries > longest_chain) {
 379                 longest_chain = entries;
 380                 longest_chain_cachesize = num_drc_entries;
 381         } else if (entries == longest_chain) {
 382                 /* prefer to keep the smallest cachesize possible here */
 383                 longest_chain_cachesize = min(longest_chain_cachesize,
 384                                                 num_drc_entries);
 385         }
 386
 387         return ret;
 388 }
 389
 390 /*
 391  * Try to find an entry matching the current call in the cache. When none
 392  * is found, we try to grab the oldest expired entry off the LRU list. If
 393  * a suitable one isn't there, then drop the cache_lock and allocate a
 394  * new one, then search again in case one got inserted while this thread
 395  * didn't hold the lock.
 396  */
 397 int
 398 nfsd_cache_lookup(struct svc_rqst *rqstp)
 399 {
 400         struct svc_cacherep     *rp, *found;
 401         __be32                  xid = rqstp->rq_xid;
 402         u32                     proto =  rqstp->rq_prot,
 403                                 vers = rqstp->rq_vers,
 404                                 proc = rqstp->rq_proc;
 405         __wsum                  csum;
 406         unsigned long           age;
 407         int type = rqstp->rq_cachetype;
 408         int rtn = RC_DOIT;
 409
 410         rqstp->rq_cacherep = NULL;
 411         if (type == RC_NOCACHE) {
 412                 nfsdstats.rcnocache++;
 413                 return rtn;
 414         }
 415
 416         csum = nfsd_cache_csum(rqstp);
 417
 418         /*
 419          * Since the common case is a cache miss followed by an insert,
 420          * preallocate an entry.
 421          */
 422         rp = nfsd_reply_cache_alloc();
 423         spin_lock(&cache_lock);
 424         if (likely(rp)) {
 425                 ++num_drc_entries;
 426                 drc_mem_usage += sizeof(*rp);
 427         }
 428
 429         /* go ahead and prune the cache */
 430         prune_cache_entries();
 431
 432         found = nfsd_cache_search(rqstp, csum);
 433         if (found) {
 434                 if (likely(rp))
 435                         nfsd_reply_cache_free_locked(rp);
 436                 rp = found;
 437                 goto found_entry;
 438         }
 439
 440         if (!rp) {
 441                 dprintk("nfsd: unable to allocate DRC entry!\n");
 442                 goto out;
 443         }
 444
 445         nfsdstats.rcmisses++;
 446         rqstp->rq_cacherep = rp;
 447         rp->c_state = RC_INPROG;
 448         rp->c_xid = xid;
 449         rp->c_proc = proc;
 450         rpc_copy_addr((struct sockaddr *)&rp->c_addr, svc_addr(rqstp));
 451         rpc_set_port((struct sockaddr *)&rp->c_addr, rpc_get_port(svc_addr(rqstp)));
 452         rp->c_prot = proto;
 453         rp->c_vers = vers;
 454         rp->c_len = rqstp->rq_arg.len;
 455         rp->c_csum = csum;
 456
 457         hash_refile(rp);
 458         lru_put_end(rp);
 459
 460         /* release any buffer */
 461         if (rp->c_type == RC_REPLBUFF) {
 462                 drc_mem_usage -= rp->c_replvec.iov_len;
 463                 kfree(rp->c_replvec.iov_base);
 464                 rp->c_replvec.iov_base = NULL;
 465         }
 466         rp->c_type = RC_NOCACHE;
 467  out:
 468         spin_unlock(&cache_lock);
 469         return rtn;
 470
 471 found_entry:
 472         nfsdstats.rchits++;
 473         /* We found a matching entry which is either in progress or done. */
 474         age = jiffies - rp->c_timestamp;
 475         lru_put_end(rp);
 476
 477         rtn = RC_DROPIT;
 478         /* Request being processed or excessive rexmits */
 479         if (rp->c_state == RC_INPROG || age < RC_DELAY)
 480                 goto out;
 481
 482         /* From the hall of fame of impractical attacks:
 483          * Is this a user who tries to snoop on the cache? */
 484         rtn = RC_DOIT;
 485         if (!rqstp->rq_secure && rp->c_secure)
 486                 goto out;
 487
 488         /* Compose RPC reply header */
 489         switch (rp->c_type) {
 490         case RC_NOCACHE:
 491                 break;
 492         case RC_REPLSTAT:
 493                 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
 494                 rtn = RC_REPLY;
 495                 break;
 496         case RC_REPLBUFF:
 497                 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
 498                         goto out;       /* should not happen */
 499                 rtn = RC_REPLY;
 500                 break;
 501         default:
 502                 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
 503                 nfsd_reply_cache_free_locked(rp);
 504         }
 505
 506         goto out;
 507 }
 508
 509 /*
 510  * Update a cache entry. This is called from nfsd_dispatch when
 511  * the procedure has been executed and the complete reply is in
 512  * rqstp->rq_res.
 513  *
 514  * We're copying around data here rather than swapping buffers because
 515  * the toplevel loop requires max-sized buffers, which would be a waste
 516  * of memory for a cache with a max reply size of 100 bytes (diropokres).
 517  *
 518  * If we should start to use different types of cache entries tailored
 519  * specifically for attrstat and fh's, we may save even more space.
 520  *
 521  * Also note that a cachetype of RC_NOCACHE can legally be passed when
 522  * nfsd failed to encode a reply that otherwise would have been cached.
 523  * In this case, nfsd_cache_update is called with statp == NULL.
 524  */
 525 void
 526 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
 527 {
 528         struct svc_cacherep *rp = rqstp->rq_cacherep;
 529         struct kvec     *resv = &rqstp->rq_res.head[0], *cachv;
 530         int             len;
 531         size_t          bufsize = 0;
 532
 533         if (!rp)
 534                 return;
 535
 536         len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
 537         len >>= 2;
 538
 539         /* Don't cache excessive amounts of data and XDR failures */
 540         if (!statp || len > (256 >> 2)) {
 541                 nfsd_reply_cache_free(rp);
 542                 return;
 543         }
 544
 545         switch (cachetype) {
 546         case RC_REPLSTAT:
 547                 if (len != 1)
 548                         printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
 549                 rp->c_replstat = *statp;
 550                 break;
 551         case RC_REPLBUFF:
 552                 cachv = &rp->c_replvec;
 553                 bufsize = len << 2;
 554                 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
 555                 if (!cachv->iov_base) {
 556                         nfsd_reply_cache_free(rp);
 557                         return;
 558                 }
 559                 cachv->iov_len = bufsize;
 560                 memcpy(cachv->iov_base, statp, bufsize);
 561                 break;
 562         case RC_NOCACHE:
 563                 nfsd_reply_cache_free(rp);
 564                 return;
 565         }
 566         spin_lock(&cache_lock);
 567         drc_mem_usage += bufsize;
 568         lru_put_end(rp);
 569         rp->c_secure = rqstp->rq_secure;
 570         rp->c_type = cachetype;
 571         rp->c_state = RC_DONE;
 572         spin_unlock(&cache_lock);
 573         return;
 574 }
 575
 576 /*
 577  * Copy cached reply to current reply buffer. Should always fit.
 578  * FIXME as reply is in a page, we should just attach the page, and
 579  * keep a refcount....
 580  */
 581 static int
 582 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
 583 {
 584         struct kvec     *vec = &rqstp->rq_res.head[0];
 585
 586         if (vec->iov_len + data->iov_len > PAGE_SIZE) {
 587                 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
 588                                 data->iov_len);
 589                 return 0;
 590         }
 591         memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
 592         vec->iov_len += data->iov_len;
 593         return 1;
 594 }
 595
 596 /*
 597  * Note that fields may be added, removed or reordered in the future. Programs
 598  * scraping this file for info should test the labels to ensure they're
 599  * getting the correct field.
 600  */
 601 static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
 602 {
 603         spin_lock(&cache_lock);
 604         seq_printf(m, "max entries:           %u\n", max_drc_entries);
 605         seq_printf(m, "num entries:           %u\n", num_drc_entries);
 606         seq_printf(m, "hash buckets:          %u\n", 1 << maskbits);
 607         seq_printf(m, "mem usage:             %u\n", drc_mem_usage);
 608         seq_printf(m, "cache hits:            %u\n", nfsdstats.rchits);
 609         seq_printf(m, "cache misses:          %u\n", nfsdstats.rcmisses);
 610         seq_printf(m, "not cached:            %u\n", nfsdstats.rcnocache);
 611         seq_printf(m, "payload misses:        %u\n", payload_misses);
 612         seq_printf(m, "longest chain len:     %u\n", longest_chain);
 613         seq_printf(m, "cachesize at longest:  %u\n", longest_chain_cachesize);
 614         spin_unlock(&cache_lock);
 615         return 0;
 616 }
 617
 618 int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
 619 {
 620         return single_open(file, nfsd_reply_cache_stats_show, NULL);
 621 }