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