2 * Request reply cache. This is currently a global cache, but this may
3 * change in the future and be a per-client cache.
5 * This code is heavily inspired by the 44BSD implementation, although
6 * it does things a bit differently.
8 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
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>
21 #define NFSDDBG_FACILITY NFSDDBG_REPCACHE
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.
28 #define TARGET_BUCKET_SIZE 64
30 static struct hlist_head * cache_hash;
31 static struct list_head lru_head;
32 static struct kmem_cache *drc_slab;
34 /* max number of entries allowed in the cache */
35 static unsigned int max_drc_entries;
37 /* number of significant bits in the hash value */
38 static unsigned int maskbits;
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
45 /* total number of entries */
46 static unsigned int num_drc_entries;
48 /* cache misses due only to checksum comparison failures */
49 static unsigned int payload_misses;
51 /* amount of memory (in bytes) currently consumed by the DRC */
52 static unsigned int drc_mem_usage;
54 /* longest hash chain seen */
55 static unsigned int longest_chain;
57 /* size of cache when we saw the longest hash chain */
58 static unsigned int longest_chain_cachesize;
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 int nfsd_reply_cache_shrink(struct shrinker *shrink,
63 struct shrink_control *sc);
65 static struct shrinker nfsd_reply_cache_shrinker = {
66 .shrink = nfsd_reply_cache_shrink,
71 * locking for the reply cache:
72 * A cache entry is "single use" if c_state == RC_INPROG
73 * Otherwise, it when accessing _prev or _next, the lock must be held.
75 static DEFINE_SPINLOCK(cache_lock);
76 static DECLARE_DELAYED_WORK(cache_cleaner, cache_cleaner_func);
79 * Put a cap on the size of the DRC based on the amount of available
80 * low memory in the machine.
92 * ...with a hard cap of 256k entries. In the worst case, each entry will be
93 * ~1k, so the above numbers should give a rough max of the amount of memory
97 nfsd_cache_size_limit(void)
100 unsigned long low_pages = totalram_pages - totalhigh_pages;
102 limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
103 return min_t(unsigned int, limit, 256*1024);
107 * Compute the number of hash buckets we need. Divide the max cachesize by
108 * the "target" max bucket size, and round up to next power of two.
111 nfsd_hashsize(unsigned int limit)
113 return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
116 static struct svc_cacherep *
117 nfsd_reply_cache_alloc(void)
119 struct svc_cacherep *rp;
121 rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
123 rp->c_state = RC_UNUSED;
124 rp->c_type = RC_NOCACHE;
125 INIT_LIST_HEAD(&rp->c_lru);
126 INIT_HLIST_NODE(&rp->c_hash);
132 nfsd_reply_cache_free_locked(struct svc_cacherep *rp)
134 if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
135 drc_mem_usage -= rp->c_replvec.iov_len;
136 kfree(rp->c_replvec.iov_base);
138 if (!hlist_unhashed(&rp->c_hash))
139 hlist_del(&rp->c_hash);
140 list_del(&rp->c_lru);
142 drc_mem_usage -= sizeof(*rp);
143 kmem_cache_free(drc_slab, rp);
147 nfsd_reply_cache_free(struct svc_cacherep *rp)
149 spin_lock(&cache_lock);
150 nfsd_reply_cache_free_locked(rp);
151 spin_unlock(&cache_lock);
154 int nfsd_reply_cache_init(void)
156 unsigned int hashsize;
158 INIT_LIST_HEAD(&lru_head);
159 max_drc_entries = nfsd_cache_size_limit();
161 hashsize = nfsd_hashsize(max_drc_entries);
162 maskbits = ilog2(hashsize);
164 register_shrinker(&nfsd_reply_cache_shrinker);
165 drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
170 cache_hash = kcalloc(hashsize, sizeof(struct hlist_head), GFP_KERNEL);
176 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
177 nfsd_reply_cache_shutdown();
181 void nfsd_reply_cache_shutdown(void)
183 struct svc_cacherep *rp;
185 unregister_shrinker(&nfsd_reply_cache_shrinker);
186 cancel_delayed_work_sync(&cache_cleaner);
188 while (!list_empty(&lru_head)) {
189 rp = list_entry(lru_head.next, struct svc_cacherep, c_lru);
190 nfsd_reply_cache_free_locked(rp);
197 kmem_cache_destroy(drc_slab);
203 * Move cache entry to end of LRU list, and queue the cleaner to run if it's
204 * not already scheduled.
207 lru_put_end(struct svc_cacherep *rp)
209 rp->c_timestamp = jiffies;
210 list_move_tail(&rp->c_lru, &lru_head);
211 schedule_delayed_work(&cache_cleaner, RC_EXPIRE);
215 * Move a cache entry from one hash list to another
218 hash_refile(struct svc_cacherep *rp)
220 hlist_del_init(&rp->c_hash);
221 hlist_add_head(&rp->c_hash, cache_hash + hash_32(rp->c_xid, maskbits));
225 * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
226 * Also prune the oldest ones when the total exceeds the max number of entries.
229 prune_cache_entries(void)
231 struct svc_cacherep *rp, *tmp;
233 list_for_each_entry_safe(rp, tmp, &lru_head, c_lru) {
235 * Don't free entries attached to calls that are still
236 * in-progress, but do keep scanning the list.
238 if (rp->c_state == RC_INPROG)
240 if (num_drc_entries <= max_drc_entries &&
241 time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
243 nfsd_reply_cache_free_locked(rp);
247 * Conditionally rearm the job. If we cleaned out the list, then
248 * cancel any pending run (since there won't be any work to do).
249 * Otherwise, we rearm the job or modify the existing one to run in
250 * RC_EXPIRE since we just ran the pruner.
252 if (list_empty(&lru_head))
253 cancel_delayed_work(&cache_cleaner);
255 mod_delayed_work(system_wq, &cache_cleaner, RC_EXPIRE);
259 cache_cleaner_func(struct work_struct *unused)
261 spin_lock(&cache_lock);
262 prune_cache_entries();
263 spin_unlock(&cache_lock);
267 nfsd_reply_cache_shrink(struct shrinker *shrink, struct shrink_control *sc)
271 spin_lock(&cache_lock);
273 prune_cache_entries();
274 num = num_drc_entries;
275 spin_unlock(&cache_lock);
281 * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
284 nfsd_cache_csum(struct svc_rqst *rqstp)
289 struct xdr_buf *buf = &rqstp->rq_arg;
290 const unsigned char *p = buf->head[0].iov_base;
291 size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
293 size_t len = min(buf->head[0].iov_len, csum_len);
295 /* rq_arg.head first */
296 csum = csum_partial(p, len, 0);
299 /* Continue into page array */
300 idx = buf->page_base / PAGE_SIZE;
301 base = buf->page_base & ~PAGE_MASK;
303 p = page_address(buf->pages[idx]) + base;
304 len = min_t(size_t, PAGE_SIZE - base, csum_len);
305 csum = csum_partial(p, len, csum);
314 nfsd_cache_match(struct svc_rqst *rqstp, __wsum csum, struct svc_cacherep *rp)
316 /* Check RPC header info first */
317 if (rqstp->rq_xid != rp->c_xid || rqstp->rq_proc != rp->c_proc ||
318 rqstp->rq_prot != rp->c_prot || rqstp->rq_vers != rp->c_vers ||
319 rqstp->rq_arg.len != rp->c_len ||
320 !rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) ||
321 rpc_get_port(svc_addr(rqstp)) != rpc_get_port((struct sockaddr *)&rp->c_addr))
324 /* compare checksum of NFS data */
325 if (csum != rp->c_csum) {
334 * Search the request hash for an entry that matches the given rqstp.
335 * Must be called with cache_lock held. Returns the found entry or
338 static struct svc_cacherep *
339 nfsd_cache_search(struct svc_rqst *rqstp, __wsum csum)
341 struct svc_cacherep *rp, *ret = NULL;
342 struct hlist_head *rh;
343 unsigned int entries = 0;
345 rh = &cache_hash[hash_32(rqstp->rq_xid, maskbits)];
346 hlist_for_each_entry(rp, rh, c_hash) {
348 if (nfsd_cache_match(rqstp, csum, rp)) {
354 /* tally hash chain length stats */
355 if (entries > longest_chain) {
356 longest_chain = entries;
357 longest_chain_cachesize = num_drc_entries;
358 } else if (entries == longest_chain) {
359 /* prefer to keep the smallest cachesize possible here */
360 longest_chain_cachesize = min(longest_chain_cachesize,
368 * Try to find an entry matching the current call in the cache. When none
369 * is found, we try to grab the oldest expired entry off the LRU list. If
370 * a suitable one isn't there, then drop the cache_lock and allocate a
371 * new one, then search again in case one got inserted while this thread
372 * didn't hold the lock.
375 nfsd_cache_lookup(struct svc_rqst *rqstp)
377 struct svc_cacherep *rp, *found;
378 __be32 xid = rqstp->rq_xid;
379 u32 proto = rqstp->rq_prot,
380 vers = rqstp->rq_vers,
381 proc = rqstp->rq_proc;
384 int type = rqstp->rq_cachetype;
387 rqstp->rq_cacherep = NULL;
388 if (type == RC_NOCACHE) {
389 nfsdstats.rcnocache++;
393 csum = nfsd_cache_csum(rqstp);
396 * Since the common case is a cache miss followed by an insert,
397 * preallocate an entry.
399 rp = nfsd_reply_cache_alloc();
400 spin_lock(&cache_lock);
403 drc_mem_usage += sizeof(*rp);
406 /* go ahead and prune the cache */
407 prune_cache_entries();
409 found = nfsd_cache_search(rqstp, csum);
412 nfsd_reply_cache_free_locked(rp);
418 dprintk("nfsd: unable to allocate DRC entry!\n");
422 nfsdstats.rcmisses++;
423 rqstp->rq_cacherep = rp;
424 rp->c_state = RC_INPROG;
427 rpc_copy_addr((struct sockaddr *)&rp->c_addr, svc_addr(rqstp));
428 rpc_set_port((struct sockaddr *)&rp->c_addr, rpc_get_port(svc_addr(rqstp)));
431 rp->c_len = rqstp->rq_arg.len;
437 /* release any buffer */
438 if (rp->c_type == RC_REPLBUFF) {
439 drc_mem_usage -= rp->c_replvec.iov_len;
440 kfree(rp->c_replvec.iov_base);
441 rp->c_replvec.iov_base = NULL;
443 rp->c_type = RC_NOCACHE;
445 spin_unlock(&cache_lock);
450 /* We found a matching entry which is either in progress or done. */
451 age = jiffies - rp->c_timestamp;
455 /* Request being processed or excessive rexmits */
456 if (rp->c_state == RC_INPROG || age < RC_DELAY)
459 /* From the hall of fame of impractical attacks:
460 * Is this a user who tries to snoop on the cache? */
462 if (!rqstp->rq_secure && rp->c_secure)
465 /* Compose RPC reply header */
466 switch (rp->c_type) {
470 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
474 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
475 goto out; /* should not happen */
479 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
480 nfsd_reply_cache_free_locked(rp);
487 * Update a cache entry. This is called from nfsd_dispatch when
488 * the procedure has been executed and the complete reply is in
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).
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.
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.
503 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
505 struct svc_cacherep *rp = rqstp->rq_cacherep;
506 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
513 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
516 /* Don't cache excessive amounts of data and XDR failures */
517 if (!statp || len > (256 >> 2)) {
518 nfsd_reply_cache_free(rp);
525 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
526 rp->c_replstat = *statp;
529 cachv = &rp->c_replvec;
531 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
532 if (!cachv->iov_base) {
533 nfsd_reply_cache_free(rp);
536 cachv->iov_len = bufsize;
537 memcpy(cachv->iov_base, statp, bufsize);
540 nfsd_reply_cache_free(rp);
543 spin_lock(&cache_lock);
544 drc_mem_usage += bufsize;
546 rp->c_secure = rqstp->rq_secure;
547 rp->c_type = cachetype;
548 rp->c_state = RC_DONE;
549 spin_unlock(&cache_lock);
554 * Copy cached reply to current reply buffer. Should always fit.
555 * FIXME as reply is in a page, we should just attach the page, and
556 * keep a refcount....
559 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
561 struct kvec *vec = &rqstp->rq_res.head[0];
563 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
564 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
568 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
569 vec->iov_len += data->iov_len;
574 * Note that fields may be added, removed or reordered in the future. Programs
575 * scraping this file for info should test the labels to ensure they're
576 * getting the correct field.
578 static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
580 spin_lock(&cache_lock);
581 seq_printf(m, "max entries: %u\n", max_drc_entries);
582 seq_printf(m, "num entries: %u\n", num_drc_entries);
583 seq_printf(m, "hash buckets: %u\n", 1 << maskbits);
584 seq_printf(m, "mem usage: %u\n", drc_mem_usage);
585 seq_printf(m, "cache hits: %u\n", nfsdstats.rchits);
586 seq_printf(m, "cache misses: %u\n", nfsdstats.rcmisses);
587 seq_printf(m, "not cached: %u\n", nfsdstats.rcnocache);
588 seq_printf(m, "payload misses: %u\n", payload_misses);
589 seq_printf(m, "longest chain len: %u\n", longest_chain);
590 seq_printf(m, "cachesize at longest: %u\n", longest_chain_cachesize);
591 spin_unlock(&cache_lock);
595 int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
597 return single_open(file, nfsd_reply_cache_stats_show, NULL);