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 nfsd_cache_entry_expired(struct svc_cacherep *rp)
227 return rp->c_state != RC_INPROG &&
228 time_after(jiffies, rp->c_timestamp + RC_EXPIRE);
232 * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
233 * Also prune the oldest ones when the total exceeds the max number of entries.
236 prune_cache_entries(void)
238 struct svc_cacherep *rp, *tmp;
240 list_for_each_entry_safe(rp, tmp, &lru_head, c_lru) {
241 if (!nfsd_cache_entry_expired(rp) &&
242 num_drc_entries <= max_drc_entries)
244 nfsd_reply_cache_free_locked(rp);
248 * Conditionally rearm the job. If we cleaned out the list, then
249 * cancel any pending run (since there won't be any work to do).
250 * Otherwise, we rearm the job or modify the existing one to run in
251 * RC_EXPIRE since we just ran the pruner.
253 if (list_empty(&lru_head))
254 cancel_delayed_work(&cache_cleaner);
256 mod_delayed_work(system_wq, &cache_cleaner, RC_EXPIRE);
260 cache_cleaner_func(struct work_struct *unused)
262 spin_lock(&cache_lock);
263 prune_cache_entries();
264 spin_unlock(&cache_lock);
268 nfsd_reply_cache_shrink(struct shrinker *shrink, struct shrink_control *sc)
272 spin_lock(&cache_lock);
274 prune_cache_entries();
275 num = num_drc_entries;
276 spin_unlock(&cache_lock);
282 * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
285 nfsd_cache_csum(struct svc_rqst *rqstp)
290 struct xdr_buf *buf = &rqstp->rq_arg;
291 const unsigned char *p = buf->head[0].iov_base;
292 size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
294 size_t len = min(buf->head[0].iov_len, csum_len);
296 /* rq_arg.head first */
297 csum = csum_partial(p, len, 0);
300 /* Continue into page array */
301 idx = buf->page_base / PAGE_SIZE;
302 base = buf->page_base & ~PAGE_MASK;
304 p = page_address(buf->pages[idx]) + base;
305 len = min_t(size_t, PAGE_SIZE - base, csum_len);
306 csum = csum_partial(p, len, csum);
315 nfsd_cache_match(struct svc_rqst *rqstp, __wsum csum, struct svc_cacherep *rp)
317 /* Check RPC header info first */
318 if (rqstp->rq_xid != rp->c_xid || rqstp->rq_proc != rp->c_proc ||
319 rqstp->rq_prot != rp->c_prot || rqstp->rq_vers != rp->c_vers ||
320 rqstp->rq_arg.len != rp->c_len ||
321 !rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) ||
322 rpc_get_port(svc_addr(rqstp)) != rpc_get_port((struct sockaddr *)&rp->c_addr))
325 /* compare checksum of NFS data */
326 if (csum != rp->c_csum) {
335 * Search the request hash for an entry that matches the given rqstp.
336 * Must be called with cache_lock held. Returns the found entry or
339 static struct svc_cacherep *
340 nfsd_cache_search(struct svc_rqst *rqstp, __wsum csum)
342 struct svc_cacherep *rp, *ret = NULL;
343 struct hlist_head *rh;
344 unsigned int entries = 0;
346 rh = &cache_hash[hash_32(rqstp->rq_xid, maskbits)];
347 hlist_for_each_entry(rp, rh, c_hash) {
349 if (nfsd_cache_match(rqstp, csum, rp)) {
355 /* tally hash chain length stats */
356 if (entries > longest_chain) {
357 longest_chain = entries;
358 longest_chain_cachesize = num_drc_entries;
359 } else if (entries == longest_chain) {
360 /* prefer to keep the smallest cachesize possible here */
361 longest_chain_cachesize = min(longest_chain_cachesize,
369 * Try to find an entry matching the current call in the cache. When none
370 * is found, we try to grab the oldest expired entry off the LRU list. If
371 * a suitable one isn't there, then drop the cache_lock and allocate a
372 * new one, then search again in case one got inserted while this thread
373 * didn't hold the lock.
376 nfsd_cache_lookup(struct svc_rqst *rqstp)
378 struct svc_cacherep *rp, *found;
379 __be32 xid = rqstp->rq_xid;
380 u32 proto = rqstp->rq_prot,
381 vers = rqstp->rq_vers,
382 proc = rqstp->rq_proc;
385 int type = rqstp->rq_cachetype;
388 rqstp->rq_cacherep = NULL;
389 if (type == RC_NOCACHE) {
390 nfsdstats.rcnocache++;
394 csum = nfsd_cache_csum(rqstp);
397 * Since the common case is a cache miss followed by an insert,
398 * preallocate an entry. First, try to reuse the first entry on the LRU
399 * if it works, then go ahead and prune the LRU list.
401 spin_lock(&cache_lock);
402 if (!list_empty(&lru_head)) {
403 rp = list_first_entry(&lru_head, struct svc_cacherep, c_lru);
404 if (nfsd_cache_entry_expired(rp) ||
405 num_drc_entries >= max_drc_entries) {
407 prune_cache_entries();
412 /* No expired ones available, allocate a new one. */
413 spin_unlock(&cache_lock);
414 rp = nfsd_reply_cache_alloc();
415 spin_lock(&cache_lock);
418 drc_mem_usage += sizeof(*rp);
422 found = nfsd_cache_search(rqstp, csum);
425 nfsd_reply_cache_free_locked(rp);
431 dprintk("nfsd: unable to allocate DRC entry!\n");
436 * We're keeping the one we just allocated. Are we now over the
437 * limit? Prune one off the tip of the LRU in trade for the one we
438 * just allocated if so.
440 if (num_drc_entries >= max_drc_entries)
441 nfsd_reply_cache_free_locked(list_first_entry(&lru_head,
442 struct svc_cacherep, c_lru));
444 nfsdstats.rcmisses++;
445 rqstp->rq_cacherep = rp;
446 rp->c_state = RC_INPROG;
449 rpc_copy_addr((struct sockaddr *)&rp->c_addr, svc_addr(rqstp));
450 rpc_set_port((struct sockaddr *)&rp->c_addr, rpc_get_port(svc_addr(rqstp)));
453 rp->c_len = rqstp->rq_arg.len;
459 /* release any buffer */
460 if (rp->c_type == RC_REPLBUFF) {
461 drc_mem_usage -= rp->c_replvec.iov_len;
462 kfree(rp->c_replvec.iov_base);
463 rp->c_replvec.iov_base = NULL;
465 rp->c_type = RC_NOCACHE;
467 spin_unlock(&cache_lock);
472 /* We found a matching entry which is either in progress or done. */
473 age = jiffies - rp->c_timestamp;
477 /* Request being processed or excessive rexmits */
478 if (rp->c_state == RC_INPROG || age < RC_DELAY)
481 /* From the hall of fame of impractical attacks:
482 * Is this a user who tries to snoop on the cache? */
484 if (!rqstp->rq_secure && rp->c_secure)
487 /* Compose RPC reply header */
488 switch (rp->c_type) {
492 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
496 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
497 goto out; /* should not happen */
501 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
502 nfsd_reply_cache_free_locked(rp);
509 * Update a cache entry. This is called from nfsd_dispatch when
510 * the procedure has been executed and the complete reply is in
513 * We're copying around data here rather than swapping buffers because
514 * the toplevel loop requires max-sized buffers, which would be a waste
515 * of memory for a cache with a max reply size of 100 bytes (diropokres).
517 * If we should start to use different types of cache entries tailored
518 * specifically for attrstat and fh's, we may save even more space.
520 * Also note that a cachetype of RC_NOCACHE can legally be passed when
521 * nfsd failed to encode a reply that otherwise would have been cached.
522 * In this case, nfsd_cache_update is called with statp == NULL.
525 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
527 struct svc_cacherep *rp = rqstp->rq_cacherep;
528 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
535 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
538 /* Don't cache excessive amounts of data and XDR failures */
539 if (!statp || len > (256 >> 2)) {
540 nfsd_reply_cache_free(rp);
547 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
548 rp->c_replstat = *statp;
551 cachv = &rp->c_replvec;
553 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
554 if (!cachv->iov_base) {
555 nfsd_reply_cache_free(rp);
558 cachv->iov_len = bufsize;
559 memcpy(cachv->iov_base, statp, bufsize);
562 nfsd_reply_cache_free(rp);
565 spin_lock(&cache_lock);
566 drc_mem_usage += bufsize;
568 rp->c_secure = rqstp->rq_secure;
569 rp->c_type = cachetype;
570 rp->c_state = RC_DONE;
571 spin_unlock(&cache_lock);
576 * Copy cached reply to current reply buffer. Should always fit.
577 * FIXME as reply is in a page, we should just attach the page, and
578 * keep a refcount....
581 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
583 struct kvec *vec = &rqstp->rq_res.head[0];
585 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
586 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
590 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
591 vec->iov_len += data->iov_len;
596 * Note that fields may be added, removed or reordered in the future. Programs
597 * scraping this file for info should test the labels to ensure they're
598 * getting the correct field.
600 static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
602 spin_lock(&cache_lock);
603 seq_printf(m, "max entries: %u\n", max_drc_entries);
604 seq_printf(m, "num entries: %u\n", num_drc_entries);
605 seq_printf(m, "hash buckets: %u\n", 1 << maskbits);
606 seq_printf(m, "mem usage: %u\n", drc_mem_usage);
607 seq_printf(m, "cache hits: %u\n", nfsdstats.rchits);
608 seq_printf(m, "cache misses: %u\n", nfsdstats.rcmisses);
609 seq_printf(m, "not cached: %u\n", nfsdstats.rcnocache);
610 seq_printf(m, "payload misses: %u\n", payload_misses);
611 seq_printf(m, "longest chain len: %u\n", longest_chain);
612 seq_printf(m, "cachesize at longest: %u\n", longest_chain_cachesize);
613 spin_unlock(&cache_lock);
617 int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
619 return single_open(file, nfsd_reply_cache_stats_show, NULL);