1 // SPDX-License-Identifier: GPL-2.0
3 * Request reply cache. This is currently a global cache, but this may
4 * change in the future and be a per-client cache.
6 * This code is heavily inspired by the 44BSD implementation, although
7 * it does things a bit differently.
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>
23 #define NFSDDBG_FACILITY NFSDDBG_REPCACHE
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.
30 #define TARGET_BUCKET_SIZE 64
32 struct nfsd_drc_bucket {
33 struct list_head lru_head;
34 spinlock_t cache_lock;
37 static struct nfsd_drc_bucket *drc_hashtbl;
38 static struct kmem_cache *drc_slab;
40 /* max number of entries allowed in the cache */
41 static unsigned int max_drc_entries;
43 /* number of significant bits in the hash value */
44 static unsigned int maskbits;
45 static unsigned int drc_hashsize;
48 * Stats and other tracking of on the duplicate reply cache. All of these and
49 * the "rc" fields in nfsdstats are protected by the cache_lock
52 /* total number of entries */
53 static atomic_t num_drc_entries;
55 /* cache misses due only to checksum comparison failures */
56 static unsigned int payload_misses;
58 /* amount of memory (in bytes) currently consumed by the DRC */
59 static unsigned int drc_mem_usage;
61 /* longest hash chain seen */
62 static unsigned int longest_chain;
64 /* size of cache when we saw the longest hash chain */
65 static unsigned int longest_chain_cachesize;
67 static int nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
68 static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
69 struct shrink_control *sc);
70 static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
71 struct shrink_control *sc);
73 static struct shrinker nfsd_reply_cache_shrinker = {
74 .scan_objects = nfsd_reply_cache_scan,
75 .count_objects = nfsd_reply_cache_count,
80 * Put a cap on the size of the DRC based on the amount of available
81 * low memory in the machine.
93 * ...with a hard cap of 256k entries. In the worst case, each entry will be
94 * ~1k, so the above numbers should give a rough max of the amount of memory
98 nfsd_cache_size_limit(void)
101 unsigned long low_pages = totalram_pages - totalhigh_pages;
103 limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
104 return min_t(unsigned int, limit, 256*1024);
108 * Compute the number of hash buckets we need. Divide the max cachesize by
109 * the "target" max bucket size, and round up to next power of two.
112 nfsd_hashsize(unsigned int limit)
114 return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
118 nfsd_cache_hash(__be32 xid)
120 return hash_32(be32_to_cpu(xid), maskbits);
123 static struct svc_cacherep *
124 nfsd_reply_cache_alloc(void)
126 struct svc_cacherep *rp;
128 rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
130 rp->c_state = RC_UNUSED;
131 rp->c_type = RC_NOCACHE;
132 INIT_LIST_HEAD(&rp->c_lru);
138 nfsd_reply_cache_free_locked(struct svc_cacherep *rp)
140 if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
141 drc_mem_usage -= rp->c_replvec.iov_len;
142 kfree(rp->c_replvec.iov_base);
144 list_del(&rp->c_lru);
145 atomic_dec(&num_drc_entries);
146 drc_mem_usage -= sizeof(*rp);
147 kmem_cache_free(drc_slab, rp);
151 nfsd_reply_cache_free(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
153 spin_lock(&b->cache_lock);
154 nfsd_reply_cache_free_locked(rp);
155 spin_unlock(&b->cache_lock);
158 int nfsd_reply_cache_init(void)
160 unsigned int hashsize;
164 max_drc_entries = nfsd_cache_size_limit();
165 atomic_set(&num_drc_entries, 0);
166 hashsize = nfsd_hashsize(max_drc_entries);
167 maskbits = ilog2(hashsize);
169 status = register_shrinker(&nfsd_reply_cache_shrinker);
173 drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
178 drc_hashtbl = kcalloc(hashsize, sizeof(*drc_hashtbl), GFP_KERNEL);
180 drc_hashtbl = vzalloc(array_size(hashsize,
181 sizeof(*drc_hashtbl)));
186 for (i = 0; i < hashsize; i++) {
187 INIT_LIST_HEAD(&drc_hashtbl[i].lru_head);
188 spin_lock_init(&drc_hashtbl[i].cache_lock);
190 drc_hashsize = hashsize;
194 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
195 nfsd_reply_cache_shutdown();
199 void nfsd_reply_cache_shutdown(void)
201 struct svc_cacherep *rp;
204 unregister_shrinker(&nfsd_reply_cache_shrinker);
206 for (i = 0; i < drc_hashsize; i++) {
207 struct list_head *head = &drc_hashtbl[i].lru_head;
208 while (!list_empty(head)) {
209 rp = list_first_entry(head, struct svc_cacherep, c_lru);
210 nfsd_reply_cache_free_locked(rp);
218 kmem_cache_destroy(drc_slab);
223 * Move cache entry to end of LRU list, and queue the cleaner to run if it's
224 * not already scheduled.
227 lru_put_end(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
229 rp->c_timestamp = jiffies;
230 list_move_tail(&rp->c_lru, &b->lru_head);
234 prune_bucket(struct nfsd_drc_bucket *b)
236 struct svc_cacherep *rp, *tmp;
239 list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
241 * Don't free entries attached to calls that are still
242 * in-progress, but do keep scanning the list.
244 if (rp->c_state == RC_INPROG)
246 if (atomic_read(&num_drc_entries) <= max_drc_entries &&
247 time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
249 nfsd_reply_cache_free_locked(rp);
256 * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
257 * Also prune the oldest ones when the total exceeds the max number of entries.
260 prune_cache_entries(void)
265 for (i = 0; i < drc_hashsize; i++) {
266 struct nfsd_drc_bucket *b = &drc_hashtbl[i];
268 if (list_empty(&b->lru_head))
270 spin_lock(&b->cache_lock);
271 freed += prune_bucket(b);
272 spin_unlock(&b->cache_lock);
278 nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
280 return atomic_read(&num_drc_entries);
284 nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
286 return prune_cache_entries();
289 * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
292 nfsd_cache_csum(struct svc_rqst *rqstp)
297 struct xdr_buf *buf = &rqstp->rq_arg;
298 const unsigned char *p = buf->head[0].iov_base;
299 size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
301 size_t len = min(buf->head[0].iov_len, csum_len);
303 /* rq_arg.head first */
304 csum = csum_partial(p, len, 0);
307 /* Continue into page array */
308 idx = buf->page_base / PAGE_SIZE;
309 base = buf->page_base & ~PAGE_MASK;
311 p = page_address(buf->pages[idx]) + base;
312 len = min_t(size_t, PAGE_SIZE - base, csum_len);
313 csum = csum_partial(p, len, csum);
322 nfsd_cache_match(struct svc_rqst *rqstp, __wsum csum, struct svc_cacherep *rp)
324 /* Check RPC XID first */
325 if (rqstp->rq_xid != rp->c_xid)
327 /* compare checksum of NFS data */
328 if (csum != rp->c_csum) {
333 /* Other discriminators */
334 if (rqstp->rq_proc != rp->c_proc ||
335 rqstp->rq_prot != rp->c_prot ||
336 rqstp->rq_vers != rp->c_vers ||
337 rqstp->rq_arg.len != rp->c_len ||
338 !rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) ||
339 rpc_get_port(svc_addr(rqstp)) != rpc_get_port((struct sockaddr *)&rp->c_addr))
346 * Search the request hash for an entry that matches the given rqstp.
347 * Must be called with cache_lock held. Returns the found entry or
350 static struct svc_cacherep *
351 nfsd_cache_search(struct nfsd_drc_bucket *b, struct svc_rqst *rqstp,
354 struct svc_cacherep *rp, *ret = NULL;
355 struct list_head *rh = &b->lru_head;
356 unsigned int entries = 0;
358 list_for_each_entry(rp, rh, c_lru) {
360 if (nfsd_cache_match(rqstp, csum, rp)) {
366 /* tally hash chain length stats */
367 if (entries > longest_chain) {
368 longest_chain = entries;
369 longest_chain_cachesize = atomic_read(&num_drc_entries);
370 } else if (entries == longest_chain) {
371 /* prefer to keep the smallest cachesize possible here */
372 longest_chain_cachesize = min_t(unsigned int,
373 longest_chain_cachesize,
374 atomic_read(&num_drc_entries));
381 * Try to find an entry matching the current call in the cache. When none
382 * is found, we try to grab the oldest expired entry off the LRU list. If
383 * a suitable one isn't there, then drop the cache_lock and allocate a
384 * new one, then search again in case one got inserted while this thread
385 * didn't hold the lock.
388 nfsd_cache_lookup(struct svc_rqst *rqstp)
390 struct svc_cacherep *rp, *found;
391 __be32 xid = rqstp->rq_xid;
392 u32 proto = rqstp->rq_prot,
393 vers = rqstp->rq_vers,
394 proc = rqstp->rq_proc;
396 u32 hash = nfsd_cache_hash(xid);
397 struct nfsd_drc_bucket *b = &drc_hashtbl[hash];
399 int type = rqstp->rq_cachetype;
402 rqstp->rq_cacherep = NULL;
403 if (type == RC_NOCACHE) {
404 nfsdstats.rcnocache++;
408 csum = nfsd_cache_csum(rqstp);
411 * Since the common case is a cache miss followed by an insert,
412 * preallocate an entry.
414 rp = nfsd_reply_cache_alloc();
415 spin_lock(&b->cache_lock);
417 atomic_inc(&num_drc_entries);
418 drc_mem_usage += sizeof(*rp);
421 /* go ahead and prune the cache */
424 found = nfsd_cache_search(b, rqstp, csum);
427 nfsd_reply_cache_free_locked(rp);
433 dprintk("nfsd: unable to allocate DRC entry!\n");
437 nfsdstats.rcmisses++;
438 rqstp->rq_cacherep = rp;
439 rp->c_state = RC_INPROG;
442 rpc_copy_addr((struct sockaddr *)&rp->c_addr, svc_addr(rqstp));
443 rpc_set_port((struct sockaddr *)&rp->c_addr, rpc_get_port(svc_addr(rqstp)));
446 rp->c_len = rqstp->rq_arg.len;
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;
457 rp->c_type = RC_NOCACHE;
459 spin_unlock(&b->cache_lock);
464 /* We found a matching entry which is either in progress or done. */
465 age = jiffies - rp->c_timestamp;
469 /* Request being processed or excessive rexmits */
470 if (rp->c_state == RC_INPROG || age < RC_DELAY)
473 /* From the hall of fame of impractical attacks:
474 * Is this a user who tries to snoop on the cache? */
476 if (!test_bit(RQ_SECURE, &rqstp->rq_flags) && rp->c_secure)
479 /* Compose RPC reply header */
480 switch (rp->c_type) {
484 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
488 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
489 goto out; /* should not happen */
493 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
494 nfsd_reply_cache_free_locked(rp);
501 * Update a cache entry. This is called from nfsd_dispatch when
502 * the procedure has been executed and the complete reply is in
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).
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.
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.
517 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
519 struct svc_cacherep *rp = rqstp->rq_cacherep;
520 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
522 struct nfsd_drc_bucket *b;
529 hash = nfsd_cache_hash(rp->c_xid);
530 b = &drc_hashtbl[hash];
532 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
535 /* Don't cache excessive amounts of data and XDR failures */
536 if (!statp || len > (256 >> 2)) {
537 nfsd_reply_cache_free(b, rp);
544 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
545 rp->c_replstat = *statp;
548 cachv = &rp->c_replvec;
550 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
551 if (!cachv->iov_base) {
552 nfsd_reply_cache_free(b, rp);
555 cachv->iov_len = bufsize;
556 memcpy(cachv->iov_base, statp, bufsize);
559 nfsd_reply_cache_free(b, rp);
562 spin_lock(&b->cache_lock);
563 drc_mem_usage += bufsize;
565 rp->c_secure = test_bit(RQ_SECURE, &rqstp->rq_flags);
566 rp->c_type = cachetype;
567 rp->c_state = RC_DONE;
568 spin_unlock(&b->cache_lock);
573 * Copy cached reply to current reply buffer. Should always fit.
574 * FIXME as reply is in a page, we should just attach the page, and
575 * keep a refcount....
578 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
580 struct kvec *vec = &rqstp->rq_res.head[0];
582 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
583 printk(KERN_WARNING "nfsd: cached reply too large (%zd).\n",
587 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
588 vec->iov_len += data->iov_len;
593 * Note that fields may be added, removed or reordered in the future. Programs
594 * scraping this file for info should test the labels to ensure they're
595 * getting the correct field.
597 static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
599 seq_printf(m, "max entries: %u\n", max_drc_entries);
600 seq_printf(m, "num entries: %u\n",
601 atomic_read(&num_drc_entries));
602 seq_printf(m, "hash buckets: %u\n", 1 << maskbits);
603 seq_printf(m, "mem usage: %u\n", drc_mem_usage);
604 seq_printf(m, "cache hits: %u\n", nfsdstats.rchits);
605 seq_printf(m, "cache misses: %u\n", nfsdstats.rcmisses);
606 seq_printf(m, "not cached: %u\n", nfsdstats.rcnocache);
607 seq_printf(m, "payload misses: %u\n", payload_misses);
608 seq_printf(m, "longest chain len: %u\n", longest_chain);
609 seq_printf(m, "cachesize at longest: %u\n", longest_chain_cachesize);
613 int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
615 return single_open(file, nfsd_reply_cache_stats_show, NULL);