fs/nfsd/nfscache.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Request reply cache. This is currently a global cache, but this may
   4  * change in the future and be a per-client cache.
   5  *
   6  * This code is heavily inspired by the 44BSD implementation, although
   7  * it does things a bit differently.
   8  *
   9  * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
  10  */
  11
  12 #include <linux/slab.h>
  13 #include <linux/vmalloc.h>
  14 #include <linux/sunrpc/addr.h>
  15 #include <linux/highmem.h>
  16 #include <linux/log2.h>
  17 #include <linux/hash.h>
  18 #include <net/checksum.h>
  19
  20 #include "nfsd.h"
  21 #include "cache.h"
  22
  23 #define NFSDDBG_FACILITY        NFSDDBG_REPCACHE
  24
  25 /*
  26  * We use this value to determine the number of hash buckets from the max
  27  * cache size, the idea being that when the cache is at its maximum number
  28  * of entries, then this should be the average number of entries per bucket.
  29  */
  30 #define TARGET_BUCKET_SIZE      64
  31
  32 struct nfsd_drc_bucket {
  33         struct rb_root rb_head;
  34         struct list_head lru_head;
  35         spinlock_t cache_lock;
  36 };
  37
  38 static struct nfsd_drc_bucket   *drc_hashtbl;
  39 static struct kmem_cache        *drc_slab;
  40
  41 /* max number of entries allowed in the cache */
  42 static unsigned int             max_drc_entries;
  43
  44 /* number of significant bits in the hash value */
  45 static unsigned int             maskbits;
  46 static unsigned int             drc_hashsize;
  47
  48 /*
  49  * Stats and other tracking of on the duplicate reply cache. All of these and
  50  * the "rc" fields in nfsdstats are protected by the cache_lock
  51  */
  52
  53 /* total number of entries */
  54 static atomic_t                 num_drc_entries;
  55
  56 /* cache misses due only to checksum comparison failures */
  57 static unsigned int             payload_misses;
  58
  59 /* amount of memory (in bytes) currently consumed by the DRC */
  60 static unsigned int             drc_mem_usage;
  61
  62 /* longest hash chain seen */
  63 static unsigned int             longest_chain;
  64
  65 /* size of cache when we saw the longest hash chain */
  66 static unsigned int             longest_chain_cachesize;
  67
  68 static int      nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
  69 static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
  70                                             struct shrink_control *sc);
  71 static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
  72                                            struct shrink_control *sc);
  73
  74 static struct shrinker nfsd_reply_cache_shrinker = {
  75         .scan_objects = nfsd_reply_cache_scan,
  76         .count_objects = nfsd_reply_cache_count,
  77         .seeks  = 1,
  78 };
  79
  80 /*
  81  * Put a cap on the size of the DRC based on the amount of available
  82  * low memory in the machine.
  83  *
  84  *  64MB:    8192
  85  * 128MB:   11585
  86  * 256MB:   16384
  87  * 512MB:   23170
  88  *   1GB:   32768
  89  *   2GB:   46340
  90  *   4GB:   65536
  91  *   8GB:   92681
  92  *  16GB:  131072
  93  *
  94  * ...with a hard cap of 256k entries. In the worst case, each entry will be
  95  * ~1k, so the above numbers should give a rough max of the amount of memory
  96  * used in k.
  97  */
  98 static unsigned int
  99 nfsd_cache_size_limit(void)
 100 {
 101         unsigned int limit;
 102         unsigned long low_pages = totalram_pages() - totalhigh_pages();
 103
 104         limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
 105         return min_t(unsigned int, limit, 256*1024);
 106 }
 107
 108 /*
 109  * Compute the number of hash buckets we need. Divide the max cachesize by
 110  * the "target" max bucket size, and round up to next power of two.
 111  */
 112 static unsigned int
 113 nfsd_hashsize(unsigned int limit)
 114 {
 115         return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
 116 }
 117
 118 static u32
 119 nfsd_cache_hash(__be32 xid)
 120 {
 121         return hash_32(be32_to_cpu(xid), maskbits);
 122 }
 123
 124 static struct svc_cacherep *
 125 nfsd_reply_cache_alloc(struct svc_rqst *rqstp, __wsum csum)
 126 {
 127         struct svc_cacherep     *rp;
 128
 129         rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
 130         if (rp) {
 131                 rp->c_state = RC_UNUSED;
 132                 rp->c_type = RC_NOCACHE;
 133                 RB_CLEAR_NODE(&rp->c_node);
 134                 INIT_LIST_HEAD(&rp->c_lru);
 135
 136                 memset(&rp->c_key, 0, sizeof(rp->c_key));
 137                 rp->c_key.k_xid = rqstp->rq_xid;
 138                 rp->c_key.k_proc = rqstp->rq_proc;
 139                 rpc_copy_addr((struct sockaddr *)&rp->c_key.k_addr, svc_addr(rqstp));
 140                 rpc_set_port((struct sockaddr *)&rp->c_key.k_addr, rpc_get_port(svc_addr(rqstp)));
 141                 rp->c_key.k_prot = rqstp->rq_prot;
 142                 rp->c_key.k_vers = rqstp->rq_vers;
 143                 rp->c_key.k_len = rqstp->rq_arg.len;
 144                 rp->c_key.k_csum = csum;
 145         }
 146         return rp;
 147 }
 148
 149 static void
 150 nfsd_reply_cache_free_locked(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
 151 {
 152         if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
 153                 drc_mem_usage -= rp->c_replvec.iov_len;
 154                 kfree(rp->c_replvec.iov_base);
 155         }
 156         if (rp->c_state != RC_UNUSED) {
 157                 rb_erase(&rp->c_node, &b->rb_head);
 158                 list_del(&rp->c_lru);
 159                 atomic_dec(&num_drc_entries);
 160                 drc_mem_usage -= sizeof(*rp);
 161         }
 162         kmem_cache_free(drc_slab, rp);
 163 }
 164
 165 static void
 166 nfsd_reply_cache_free(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
 167 {
 168         spin_lock(&b->cache_lock);
 169         nfsd_reply_cache_free_locked(b, rp);
 170         spin_unlock(&b->cache_lock);
 171 }
 172
 173 int nfsd_reply_cache_init(void)
 174 {
 175         unsigned int hashsize;
 176         unsigned int i;
 177         int status = 0;
 178
 179         max_drc_entries = nfsd_cache_size_limit();
 180         atomic_set(&num_drc_entries, 0);
 181         hashsize = nfsd_hashsize(max_drc_entries);
 182         maskbits = ilog2(hashsize);
 183
 184         status = register_shrinker(&nfsd_reply_cache_shrinker);
 185         if (status)
 186                 return status;
 187
 188         drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
 189                                         0, 0, NULL);
 190         if (!drc_slab)
 191                 goto out_nomem;
 192
 193         drc_hashtbl = kcalloc(hashsize, sizeof(*drc_hashtbl), GFP_KERNEL);
 194         if (!drc_hashtbl) {
 195                 drc_hashtbl = vzalloc(array_size(hashsize,
 196                                                  sizeof(*drc_hashtbl)));
 197                 if (!drc_hashtbl)
 198                         goto out_nomem;
 199         }
 200
 201         for (i = 0; i < hashsize; i++) {
 202                 INIT_LIST_HEAD(&drc_hashtbl[i].lru_head);
 203                 spin_lock_init(&drc_hashtbl[i].cache_lock);
 204         }
 205         drc_hashsize = hashsize;
 206
 207         return 0;
 208 out_nomem:
 209         printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
 210         nfsd_reply_cache_shutdown();
 211         return -ENOMEM;
 212 }
 213
 214 void nfsd_reply_cache_shutdown(void)
 215 {
 216         struct svc_cacherep     *rp;
 217         unsigned int i;
 218
 219         unregister_shrinker(&nfsd_reply_cache_shrinker);
 220
 221         for (i = 0; i < drc_hashsize; i++) {
 222                 struct list_head *head = &drc_hashtbl[i].lru_head;
 223                 while (!list_empty(head)) {
 224                         rp = list_first_entry(head, struct svc_cacherep, c_lru);
 225                         nfsd_reply_cache_free_locked(&drc_hashtbl[i], rp);
 226                 }
 227         }
 228
 229         kvfree(drc_hashtbl);
 230         drc_hashtbl = NULL;
 231         drc_hashsize = 0;
 232
 233         kmem_cache_destroy(drc_slab);
 234         drc_slab = NULL;
 235 }
 236
 237 /*
 238  * Move cache entry to end of LRU list, and queue the cleaner to run if it's
 239  * not already scheduled.
 240  */
 241 static void
 242 lru_put_end(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
 243 {
 244         rp->c_timestamp = jiffies;
 245         list_move_tail(&rp->c_lru, &b->lru_head);
 246 }
 247
 248 static long
 249 prune_bucket(struct nfsd_drc_bucket *b)
 250 {
 251         struct svc_cacherep *rp, *tmp;
 252         long freed = 0;
 253
 254         list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
 255                 /*
 256                  * Don't free entries attached to calls that are still
 257                  * in-progress, but do keep scanning the list.
 258                  */
 259                 if (rp->c_state == RC_INPROG)
 260                         continue;
 261                 if (atomic_read(&num_drc_entries) <= max_drc_entries &&
 262                     time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
 263                         break;
 264                 nfsd_reply_cache_free_locked(b, rp);
 265                 freed++;
 266         }
 267         return freed;
 268 }
 269
 270 /*
 271  * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
 272  * Also prune the oldest ones when the total exceeds the max number of entries.
 273  */
 274 static long
 275 prune_cache_entries(void)
 276 {
 277         unsigned int i;
 278         long freed = 0;
 279
 280         for (i = 0; i < drc_hashsize; i++) {
 281                 struct nfsd_drc_bucket *b = &drc_hashtbl[i];
 282
 283                 if (list_empty(&b->lru_head))
 284                         continue;
 285                 spin_lock(&b->cache_lock);
 286                 freed += prune_bucket(b);
 287                 spin_unlock(&b->cache_lock);
 288         }
 289         return freed;
 290 }
 291
 292 static unsigned long
 293 nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
 294 {
 295         return atomic_read(&num_drc_entries);
 296 }
 297
 298 static unsigned long
 299 nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
 300 {
 301         return prune_cache_entries();
 302 }
 303 /*
 304  * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
 305  */
 306 static __wsum
 307 nfsd_cache_csum(struct svc_rqst *rqstp)
 308 {
 309         int idx;
 310         unsigned int base;
 311         __wsum csum;
 312         struct xdr_buf *buf = &rqstp->rq_arg;
 313         const unsigned char *p = buf->head[0].iov_base;
 314         size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
 315                                 RC_CSUMLEN);
 316         size_t len = min(buf->head[0].iov_len, csum_len);
 317
 318         /* rq_arg.head first */
 319         csum = csum_partial(p, len, 0);
 320         csum_len -= len;
 321
 322         /* Continue into page array */
 323         idx = buf->page_base / PAGE_SIZE;
 324         base = buf->page_base & ~PAGE_MASK;
 325         while (csum_len) {
 326                 p = page_address(buf->pages[idx]) + base;
 327                 len = min_t(size_t, PAGE_SIZE - base, csum_len);
 328                 csum = csum_partial(p, len, csum);
 329                 csum_len -= len;
 330                 base = 0;
 331                 ++idx;
 332         }
 333         return csum;
 334 }
 335
 336 static int
 337 nfsd_cache_key_cmp(const struct svc_cacherep *key, const struct svc_cacherep *rp)
 338 {
 339         if (key->c_key.k_xid == rp->c_key.k_xid &&
 340             key->c_key.k_csum != rp->c_key.k_csum)
 341                 ++payload_misses;
 342
 343         return memcmp(&key->c_key, &rp->c_key, sizeof(key->c_key));
 344 }
 345
 346 /*
 347  * Search the request hash for an entry that matches the given rqstp.
 348  * Must be called with cache_lock held. Returns the found entry or
 349  * inserts an empty key on failure.
 350  */
 351 static struct svc_cacherep *
 352 nfsd_cache_insert(struct nfsd_drc_bucket *b, struct svc_cacherep *key)
 353 {
 354         struct svc_cacherep     *rp, *ret = key;
 355         struct rb_node          **p = &b->rb_head.rb_node,
 356                                 *parent = NULL;
 357         unsigned int            entries = 0;
 358         int cmp;
 359
 360         while (*p != NULL) {
 361                 ++entries;
 362                 parent = *p;
 363                 rp = rb_entry(parent, struct svc_cacherep, c_node);
 364
 365                 cmp = nfsd_cache_key_cmp(key, rp);
 366                 if (cmp < 0)
 367                         p = &parent->rb_left;
 368                 else if (cmp > 0)
 369                         p = &parent->rb_right;
 370                 else {
 371                         ret = rp;
 372                         goto out;
 373                 }
 374         }
 375         rb_link_node(&key->c_node, parent, p);
 376         rb_insert_color(&key->c_node, &b->rb_head);
 377 out:
 378         /* tally hash chain length stats */
 379         if (entries > longest_chain) {
 380                 longest_chain = entries;
 381                 longest_chain_cachesize = atomic_read(&num_drc_entries);
 382         } else if (entries == longest_chain) {
 383                 /* prefer to keep the smallest cachesize possible here */
 384                 longest_chain_cachesize = min_t(unsigned int,
 385                                 longest_chain_cachesize,
 386                                 atomic_read(&num_drc_entries));
 387         }
 388
 389         lru_put_end(b, ret);
 390         return ret;
 391 }
 392
 393 /*
 394  * Try to find an entry matching the current call in the cache. When none
 395  * is found, we try to grab the oldest expired entry off the LRU list. If
 396  * a suitable one isn't there, then drop the cache_lock and allocate a
 397  * new one, then search again in case one got inserted while this thread
 398  * didn't hold the lock.
 399  */
 400 int
 401 nfsd_cache_lookup(struct svc_rqst *rqstp)
 402 {
 403         struct svc_cacherep     *rp, *found;
 404         __be32                  xid = rqstp->rq_xid;
 405         __wsum                  csum;
 406         u32 hash = nfsd_cache_hash(xid);
 407         struct nfsd_drc_bucket *b = &drc_hashtbl[hash];
 408         int type = rqstp->rq_cachetype;
 409         int rtn = RC_DOIT;
 410
 411         rqstp->rq_cacherep = NULL;
 412         if (type == RC_NOCACHE) {
 413                 nfsdstats.rcnocache++;
 414                 return rtn;
 415         }
 416
 417         csum = nfsd_cache_csum(rqstp);
 418
 419         /*
 420          * Since the common case is a cache miss followed by an insert,
 421          * preallocate an entry.
 422          */
 423         rp = nfsd_reply_cache_alloc(rqstp, csum);
 424         if (!rp) {
 425                 dprintk("nfsd: unable to allocate DRC entry!\n");
 426                 return rtn;
 427         }
 428
 429         spin_lock(&b->cache_lock);
 430         found = nfsd_cache_insert(b, rp);
 431         if (found != rp) {
 432                 nfsd_reply_cache_free_locked(NULL, rp);
 433                 rp = found;
 434                 goto found_entry;
 435         }
 436
 437         nfsdstats.rcmisses++;
 438         rqstp->rq_cacherep = rp;
 439         rp->c_state = RC_INPROG;
 440
 441         atomic_inc(&num_drc_entries);
 442         drc_mem_usage += sizeof(*rp);
 443
 444         /* go ahead and prune the cache */
 445         prune_bucket(b);
 446  out:
 447         spin_unlock(&b->cache_lock);
 448         return rtn;
 449
 450 found_entry:
 451         /* We found a matching entry which is either in progress or done. */
 452         nfsdstats.rchits++;
 453         rtn = RC_DROPIT;
 454
 455         /* Request being processed */
 456         if (rp->c_state == RC_INPROG)
 457                 goto out;
 458
 459         /* From the hall of fame of impractical attacks:
 460          * Is this a user who tries to snoop on the cache? */
 461         rtn = RC_DOIT;
 462         if (!test_bit(RQ_SECURE, &rqstp->rq_flags) && rp->c_secure)
 463                 goto out;
 464
 465         /* Compose RPC reply header */
 466         switch (rp->c_type) {
 467         case RC_NOCACHE:
 468                 break;
 469         case RC_REPLSTAT:
 470                 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
 471                 rtn = RC_REPLY;
 472                 break;
 473         case RC_REPLBUFF:
 474                 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
 475                         goto out;       /* should not happen */
 476                 rtn = RC_REPLY;
 477                 break;
 478         default:
 479                 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
 480                 nfsd_reply_cache_free_locked(b, rp);
 481         }
 482
 483         goto out;
 484 }
 485
 486 /*
 487  * Update a cache entry. This is called from nfsd_dispatch when
 488  * the procedure has been executed and the complete reply is in
 489  * rqstp->rq_res.
 490  *
 491  * We're copying around data here rather than swapping buffers because
 492  * the toplevel loop requires max-sized buffers, which would be a waste
 493  * of memory for a cache with a max reply size of 100 bytes (diropokres).
 494  *
 495  * If we should start to use different types of cache entries tailored
 496  * specifically for attrstat and fh's, we may save even more space.
 497  *
 498  * Also note that a cachetype of RC_NOCACHE can legally be passed when
 499  * nfsd failed to encode a reply that otherwise would have been cached.
 500  * In this case, nfsd_cache_update is called with statp == NULL.
 501  */
 502 void
 503 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
 504 {
 505         struct svc_cacherep *rp = rqstp->rq_cacherep;
 506         struct kvec     *resv = &rqstp->rq_res.head[0], *cachv;
 507         u32             hash;
 508         struct nfsd_drc_bucket *b;
 509         int             len;
 510         size_t          bufsize = 0;
 511
 512         if (!rp)
 513                 return;
 514
 515         hash = nfsd_cache_hash(rp->c_key.k_xid);
 516         b = &drc_hashtbl[hash];
 517
 518         len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
 519         len >>= 2;
 520
 521         /* Don't cache excessive amounts of data and XDR failures */
 522         if (!statp || len > (256 >> 2)) {
 523                 nfsd_reply_cache_free(b, rp);
 524                 return;
 525         }
 526
 527         switch (cachetype) {
 528         case RC_REPLSTAT:
 529                 if (len != 1)
 530                         printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
 531                 rp->c_replstat = *statp;
 532                 break;
 533         case RC_REPLBUFF:
 534                 cachv = &rp->c_replvec;
 535                 bufsize = len << 2;
 536                 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
 537                 if (!cachv->iov_base) {
 538                         nfsd_reply_cache_free(b, rp);
 539                         return;
 540                 }
 541                 cachv->iov_len = bufsize;
 542                 memcpy(cachv->iov_base, statp, bufsize);
 543                 break;
 544         case RC_NOCACHE:
 545                 nfsd_reply_cache_free(b, rp);
 546                 return;
 547         }
 548         spin_lock(&b->cache_lock);
 549         drc_mem_usage += bufsize;
 550         lru_put_end(b, rp);
 551         rp->c_secure = test_bit(RQ_SECURE, &rqstp->rq_flags);
 552         rp->c_type = cachetype;
 553         rp->c_state = RC_DONE;
 554         spin_unlock(&b->cache_lock);
 555         return;
 556 }
 557
 558 /*
 559  * Copy cached reply to current reply buffer. Should always fit.
 560  * FIXME as reply is in a page, we should just attach the page, and
 561  * keep a refcount....
 562  */
 563 static int
 564 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
 565 {
 566         struct kvec     *vec = &rqstp->rq_res.head[0];
 567
 568         if (vec->iov_len + data->iov_len > PAGE_SIZE) {
 569                 printk(KERN_WARNING "nfsd: cached reply too large (%zd).\n",
 570                                 data->iov_len);
 571                 return 0;
 572         }
 573         memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
 574         vec->iov_len += data->iov_len;
 575         return 1;
 576 }
 577
 578 /*
 579  * Note that fields may be added, removed or reordered in the future. Programs
 580  * scraping this file for info should test the labels to ensure they're
 581  * getting the correct field.
 582  */
 583 static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
 584 {
 585         seq_printf(m, "max entries:           %u\n", max_drc_entries);
 586         seq_printf(m, "num entries:           %u\n",
 587                         atomic_read(&num_drc_entries));
 588         seq_printf(m, "hash buckets:          %u\n", 1 << maskbits);
 589         seq_printf(m, "mem usage:             %u\n", drc_mem_usage);
 590         seq_printf(m, "cache hits:            %u\n", nfsdstats.rchits);
 591         seq_printf(m, "cache misses:          %u\n", nfsdstats.rcmisses);
 592         seq_printf(m, "not cached:            %u\n", nfsdstats.rcnocache);
 593         seq_printf(m, "payload misses:        %u\n", payload_misses);
 594         seq_printf(m, "longest chain len:     %u\n", longest_chain);
 595         seq_printf(m, "cachesize at longest:  %u\n", longest_chain_cachesize);
 596         return 0;
 597 }
 598
 599 int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
 600 {
 601         return single_open(file, nfsd_reply_cache_stats_show, NULL);
 602 }