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 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);
67 static struct shrinker nfsd_reply_cache_shrinker = {
68 .scan_objects = nfsd_reply_cache_scan,
69 .count_objects = nfsd_reply_cache_count,
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.
78 static DEFINE_SPINLOCK(cache_lock);
79 static DECLARE_DELAYED_WORK(cache_cleaner, cache_cleaner_func);
82 * Put a cap on the size of the DRC based on the amount of available
83 * low memory in the machine.
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
100 nfsd_cache_size_limit(void)
103 unsigned long low_pages = totalram_pages - totalhigh_pages;
105 limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
106 return min_t(unsigned int, limit, 256*1024);
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.
114 nfsd_hashsize(unsigned int limit)
116 return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
119 static struct svc_cacherep *
120 nfsd_reply_cache_alloc(void)
122 struct svc_cacherep *rp;
124 rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
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);
135 nfsd_reply_cache_unhash(struct svc_cacherep *rp)
137 hlist_del_init(&rp->c_hash);
138 list_del_init(&rp->c_lru);
142 nfsd_reply_cache_free_locked(struct svc_cacherep *rp)
144 if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
145 drc_mem_usage -= rp->c_replvec.iov_len;
146 kfree(rp->c_replvec.iov_base);
148 if (!hlist_unhashed(&rp->c_hash))
149 hlist_del(&rp->c_hash);
150 list_del(&rp->c_lru);
152 drc_mem_usage -= sizeof(*rp);
153 kmem_cache_free(drc_slab, rp);
157 nfsd_reply_cache_free(struct svc_cacherep *rp)
159 spin_lock(&cache_lock);
160 nfsd_reply_cache_free_locked(rp);
161 spin_unlock(&cache_lock);
164 int nfsd_reply_cache_init(void)
166 unsigned int hashsize;
168 INIT_LIST_HEAD(&lru_head);
169 max_drc_entries = nfsd_cache_size_limit();
171 hashsize = nfsd_hashsize(max_drc_entries);
172 maskbits = ilog2(hashsize);
174 register_shrinker(&nfsd_reply_cache_shrinker);
175 drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
180 cache_hash = kcalloc(hashsize, sizeof(struct hlist_head), GFP_KERNEL);
186 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
187 nfsd_reply_cache_shutdown();
191 void nfsd_reply_cache_shutdown(void)
193 struct svc_cacherep *rp;
195 unregister_shrinker(&nfsd_reply_cache_shrinker);
196 cancel_delayed_work_sync(&cache_cleaner);
198 while (!list_empty(&lru_head)) {
199 rp = list_entry(lru_head.next, struct svc_cacherep, c_lru);
200 nfsd_reply_cache_free_locked(rp);
207 kmem_cache_destroy(drc_slab);
213 * Move cache entry to end of LRU list, and queue the cleaner to run if it's
214 * not already scheduled.
217 lru_put_end(struct svc_cacherep *rp)
219 rp->c_timestamp = jiffies;
220 list_move_tail(&rp->c_lru, &lru_head);
221 schedule_delayed_work(&cache_cleaner, RC_EXPIRE);
225 * Move a cache entry from one hash list to another
228 hash_refile(struct svc_cacherep *rp)
230 hlist_del_init(&rp->c_hash);
231 hlist_add_head(&rp->c_hash, cache_hash + hash_32(rp->c_xid, maskbits));
235 nfsd_cache_entry_expired(struct svc_cacherep *rp)
237 return rp->c_state != RC_INPROG &&
238 time_after(jiffies, rp->c_timestamp + RC_EXPIRE);
242 * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
243 * Also prune the oldest ones when the total exceeds the max number of entries.
246 prune_cache_entries(void)
248 struct svc_cacherep *rp, *tmp;
251 list_for_each_entry_safe(rp, tmp, &lru_head, c_lru) {
252 if (!nfsd_cache_entry_expired(rp) &&
253 num_drc_entries <= max_drc_entries)
255 nfsd_reply_cache_free_locked(rp);
260 * Conditionally rearm the job. If we cleaned out the list, then
261 * cancel any pending run (since there won't be any work to do).
262 * Otherwise, we rearm the job or modify the existing one to run in
263 * RC_EXPIRE since we just ran the pruner.
265 if (list_empty(&lru_head))
266 cancel_delayed_work(&cache_cleaner);
268 mod_delayed_work(system_wq, &cache_cleaner, RC_EXPIRE);
273 cache_cleaner_func(struct work_struct *unused)
275 spin_lock(&cache_lock);
276 prune_cache_entries();
277 spin_unlock(&cache_lock);
281 nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
285 spin_lock(&cache_lock);
286 num = num_drc_entries;
287 spin_unlock(&cache_lock);
293 nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
297 spin_lock(&cache_lock);
298 freed = prune_cache_entries();
299 spin_unlock(&cache_lock);
303 * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
306 nfsd_cache_csum(struct svc_rqst *rqstp)
311 struct xdr_buf *buf = &rqstp->rq_arg;
312 const unsigned char *p = buf->head[0].iov_base;
313 size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
315 size_t len = min(buf->head[0].iov_len, csum_len);
317 /* rq_arg.head first */
318 csum = csum_partial(p, len, 0);
321 /* Continue into page array */
322 idx = buf->page_base / PAGE_SIZE;
323 base = buf->page_base & ~PAGE_MASK;
325 p = page_address(buf->pages[idx]) + base;
326 len = min_t(size_t, PAGE_SIZE - base, csum_len);
327 csum = csum_partial(p, len, csum);
336 nfsd_cache_match(struct svc_rqst *rqstp, __wsum csum, struct svc_cacherep *rp)
338 /* Check RPC header info first */
339 if (rqstp->rq_xid != rp->c_xid || rqstp->rq_proc != rp->c_proc ||
340 rqstp->rq_prot != rp->c_prot || rqstp->rq_vers != rp->c_vers ||
341 rqstp->rq_arg.len != rp->c_len ||
342 !rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) ||
343 rpc_get_port(svc_addr(rqstp)) != rpc_get_port((struct sockaddr *)&rp->c_addr))
346 /* compare checksum of NFS data */
347 if (csum != rp->c_csum) {
356 * Search the request hash for an entry that matches the given rqstp.
357 * Must be called with cache_lock held. Returns the found entry or
360 static struct svc_cacherep *
361 nfsd_cache_search(struct svc_rqst *rqstp, __wsum csum)
363 struct svc_cacherep *rp, *ret = NULL;
364 struct hlist_head *rh;
365 unsigned int entries = 0;
367 rh = &cache_hash[hash_32(rqstp->rq_xid, maskbits)];
368 hlist_for_each_entry(rp, rh, c_hash) {
370 if (nfsd_cache_match(rqstp, csum, rp)) {
376 /* tally hash chain length stats */
377 if (entries > longest_chain) {
378 longest_chain = entries;
379 longest_chain_cachesize = num_drc_entries;
380 } else if (entries == longest_chain) {
381 /* prefer to keep the smallest cachesize possible here */
382 longest_chain_cachesize = min(longest_chain_cachesize,
390 * Try to find an entry matching the current call in the cache. When none
391 * is found, we try to grab the oldest expired entry off the LRU list. If
392 * a suitable one isn't there, then drop the cache_lock and allocate a
393 * new one, then search again in case one got inserted while this thread
394 * didn't hold the lock.
397 nfsd_cache_lookup(struct svc_rqst *rqstp)
399 struct svc_cacherep *rp, *found;
400 __be32 xid = rqstp->rq_xid;
401 u32 proto = rqstp->rq_prot,
402 vers = rqstp->rq_vers,
403 proc = rqstp->rq_proc;
406 int type = rqstp->rq_cachetype;
409 rqstp->rq_cacherep = NULL;
410 if (type == RC_NOCACHE) {
411 nfsdstats.rcnocache++;
415 csum = nfsd_cache_csum(rqstp);
418 * Since the common case is a cache miss followed by an insert,
419 * preallocate an entry. First, try to reuse the first entry on the LRU
420 * if it works, then go ahead and prune the LRU list.
422 spin_lock(&cache_lock);
423 if (!list_empty(&lru_head)) {
424 rp = list_first_entry(&lru_head, struct svc_cacherep, c_lru);
425 if (nfsd_cache_entry_expired(rp) ||
426 num_drc_entries >= max_drc_entries) {
427 nfsd_reply_cache_unhash(rp);
428 prune_cache_entries();
433 /* No expired ones available, allocate a new one. */
434 spin_unlock(&cache_lock);
435 rp = nfsd_reply_cache_alloc();
436 spin_lock(&cache_lock);
439 drc_mem_usage += sizeof(*rp);
443 found = nfsd_cache_search(rqstp, csum);
446 nfsd_reply_cache_free_locked(rp);
452 dprintk("nfsd: unable to allocate DRC entry!\n");
457 * We're keeping the one we just allocated. Are we now over the
458 * limit? Prune one off the tip of the LRU in trade for the one we
459 * just allocated if so.
461 if (num_drc_entries >= max_drc_entries)
462 nfsd_reply_cache_free_locked(list_first_entry(&lru_head,
463 struct svc_cacherep, c_lru));
465 nfsdstats.rcmisses++;
466 rqstp->rq_cacherep = rp;
467 rp->c_state = RC_INPROG;
470 rpc_copy_addr((struct sockaddr *)&rp->c_addr, svc_addr(rqstp));
471 rpc_set_port((struct sockaddr *)&rp->c_addr, rpc_get_port(svc_addr(rqstp)));
474 rp->c_len = rqstp->rq_arg.len;
480 /* release any buffer */
481 if (rp->c_type == RC_REPLBUFF) {
482 drc_mem_usage -= rp->c_replvec.iov_len;
483 kfree(rp->c_replvec.iov_base);
484 rp->c_replvec.iov_base = NULL;
486 rp->c_type = RC_NOCACHE;
488 spin_unlock(&cache_lock);
493 /* We found a matching entry which is either in progress or done. */
494 age = jiffies - rp->c_timestamp;
498 /* Request being processed or excessive rexmits */
499 if (rp->c_state == RC_INPROG || age < RC_DELAY)
502 /* From the hall of fame of impractical attacks:
503 * Is this a user who tries to snoop on the cache? */
505 if (!rqstp->rq_secure && rp->c_secure)
508 /* Compose RPC reply header */
509 switch (rp->c_type) {
513 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
517 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
518 goto out; /* should not happen */
522 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
523 nfsd_reply_cache_free_locked(rp);
530 * Update a cache entry. This is called from nfsd_dispatch when
531 * the procedure has been executed and the complete reply is in
534 * We're copying around data here rather than swapping buffers because
535 * the toplevel loop requires max-sized buffers, which would be a waste
536 * of memory for a cache with a max reply size of 100 bytes (diropokres).
538 * If we should start to use different types of cache entries tailored
539 * specifically for attrstat and fh's, we may save even more space.
541 * Also note that a cachetype of RC_NOCACHE can legally be passed when
542 * nfsd failed to encode a reply that otherwise would have been cached.
543 * In this case, nfsd_cache_update is called with statp == NULL.
546 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
548 struct svc_cacherep *rp = rqstp->rq_cacherep;
549 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
556 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
559 /* Don't cache excessive amounts of data and XDR failures */
560 if (!statp || len > (256 >> 2)) {
561 nfsd_reply_cache_free(rp);
568 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
569 rp->c_replstat = *statp;
572 cachv = &rp->c_replvec;
574 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
575 if (!cachv->iov_base) {
576 nfsd_reply_cache_free(rp);
579 cachv->iov_len = bufsize;
580 memcpy(cachv->iov_base, statp, bufsize);
583 nfsd_reply_cache_free(rp);
586 spin_lock(&cache_lock);
587 drc_mem_usage += bufsize;
589 rp->c_secure = rqstp->rq_secure;
590 rp->c_type = cachetype;
591 rp->c_state = RC_DONE;
592 spin_unlock(&cache_lock);
597 * Copy cached reply to current reply buffer. Should always fit.
598 * FIXME as reply is in a page, we should just attach the page, and
599 * keep a refcount....
602 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
604 struct kvec *vec = &rqstp->rq_res.head[0];
606 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
607 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
611 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
612 vec->iov_len += data->iov_len;
617 * Note that fields may be added, removed or reordered in the future. Programs
618 * scraping this file for info should test the labels to ensure they're
619 * getting the correct field.
621 static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
623 spin_lock(&cache_lock);
624 seq_printf(m, "max entries: %u\n", max_drc_entries);
625 seq_printf(m, "num entries: %u\n", num_drc_entries);
626 seq_printf(m, "hash buckets: %u\n", 1 << maskbits);
627 seq_printf(m, "mem usage: %u\n", drc_mem_usage);
628 seq_printf(m, "cache hits: %u\n", nfsdstats.rchits);
629 seq_printf(m, "cache misses: %u\n", nfsdstats.rcmisses);
630 seq_printf(m, "not cached: %u\n", nfsdstats.rcnocache);
631 seq_printf(m, "payload misses: %u\n", payload_misses);
632 seq_printf(m, "longest chain len: %u\n", longest_chain);
633 seq_printf(m, "cachesize at longest: %u\n", longest_chain_cachesize);
634 spin_unlock(&cache_lock);
638 int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
640 return single_open(file, nfsd_reply_cache_stats_show, NULL);