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.
9 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
12 #include <linux/sunrpc/svc_xprt.h>
13 #include <linux/slab.h>
14 #include <linux/vmalloc.h>
15 #include <linux/sunrpc/addr.h>
16 #include <linux/highmem.h>
17 #include <linux/log2.h>
18 #include <linux/hash.h>
19 #include <net/checksum.h>
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 rb_root rb_head;
34 struct list_head lru_head;
35 spinlock_t cache_lock;
38 static struct kmem_cache *drc_slab;
40 static int nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
41 static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
42 struct shrink_control *sc);
43 static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
44 struct shrink_control *sc);
47 * Put a cap on the size of the DRC based on the amount of available
48 * low memory in the machine.
60 * ...with a hard cap of 256k entries. In the worst case, each entry will be
61 * ~1k, so the above numbers should give a rough max of the amount of memory
64 * XXX: these limits are per-container, so memory used will increase
65 * linearly with number of containers. Maybe that's OK.
68 nfsd_cache_size_limit(void)
71 unsigned long low_pages = totalram_pages() - totalhigh_pages();
73 limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
74 return min_t(unsigned int, limit, 256*1024);
78 * Compute the number of hash buckets we need. Divide the max cachesize by
79 * the "target" max bucket size, and round up to next power of two.
82 nfsd_hashsize(unsigned int limit)
84 return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
87 static struct nfsd_cacherep *
88 nfsd_cacherep_alloc(struct svc_rqst *rqstp, __wsum csum,
91 struct nfsd_cacherep *rp;
93 rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
95 rp->c_state = RC_UNUSED;
96 rp->c_type = RC_NOCACHE;
97 RB_CLEAR_NODE(&rp->c_node);
98 INIT_LIST_HEAD(&rp->c_lru);
100 memset(&rp->c_key, 0, sizeof(rp->c_key));
101 rp->c_key.k_xid = rqstp->rq_xid;
102 rp->c_key.k_proc = rqstp->rq_proc;
103 rpc_copy_addr((struct sockaddr *)&rp->c_key.k_addr, svc_addr(rqstp));
104 rpc_set_port((struct sockaddr *)&rp->c_key.k_addr, rpc_get_port(svc_addr(rqstp)));
105 rp->c_key.k_prot = rqstp->rq_prot;
106 rp->c_key.k_vers = rqstp->rq_vers;
107 rp->c_key.k_len = rqstp->rq_arg.len;
108 rp->c_key.k_csum = csum;
113 static void nfsd_cacherep_free(struct nfsd_cacherep *rp)
115 if (rp->c_type == RC_REPLBUFF)
116 kfree(rp->c_replvec.iov_base);
117 kmem_cache_free(drc_slab, rp);
121 nfsd_cacherep_dispose(struct list_head *dispose)
123 struct nfsd_cacherep *rp;
124 unsigned long freed = 0;
126 while (!list_empty(dispose)) {
127 rp = list_first_entry(dispose, struct nfsd_cacherep, c_lru);
128 list_del(&rp->c_lru);
129 nfsd_cacherep_free(rp);
136 nfsd_cacherep_unlink_locked(struct nfsd_net *nn, struct nfsd_drc_bucket *b,
137 struct nfsd_cacherep *rp)
139 if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base)
140 nfsd_stats_drc_mem_usage_sub(nn, rp->c_replvec.iov_len);
141 if (rp->c_state != RC_UNUSED) {
142 rb_erase(&rp->c_node, &b->rb_head);
143 list_del(&rp->c_lru);
144 atomic_dec(&nn->num_drc_entries);
145 nfsd_stats_drc_mem_usage_sub(nn, sizeof(*rp));
150 nfsd_reply_cache_free_locked(struct nfsd_drc_bucket *b, struct nfsd_cacherep *rp,
153 nfsd_cacherep_unlink_locked(nn, b, rp);
154 nfsd_cacherep_free(rp);
158 nfsd_reply_cache_free(struct nfsd_drc_bucket *b, struct nfsd_cacherep *rp,
161 spin_lock(&b->cache_lock);
162 nfsd_cacherep_unlink_locked(nn, b, rp);
163 spin_unlock(&b->cache_lock);
164 nfsd_cacherep_free(rp);
167 int nfsd_drc_slab_create(void)
169 drc_slab = kmem_cache_create("nfsd_drc",
170 sizeof(struct nfsd_cacherep), 0, 0, NULL);
171 return drc_slab ? 0: -ENOMEM;
174 void nfsd_drc_slab_free(void)
176 kmem_cache_destroy(drc_slab);
180 * nfsd_net_reply_cache_init - per net namespace reply cache set-up
181 * @nn: nfsd_net being initialized
183 * Returns zero on succes; otherwise a negative errno is returned.
185 int nfsd_net_reply_cache_init(struct nfsd_net *nn)
187 return nfsd_percpu_counters_init(nn->counter, NFSD_NET_COUNTERS_NUM);
191 * nfsd_net_reply_cache_destroy - per net namespace reply cache tear-down
192 * @nn: nfsd_net being freed
195 void nfsd_net_reply_cache_destroy(struct nfsd_net *nn)
197 nfsd_percpu_counters_destroy(nn->counter, NFSD_NET_COUNTERS_NUM);
200 int nfsd_reply_cache_init(struct nfsd_net *nn)
202 unsigned int hashsize;
206 nn->max_drc_entries = nfsd_cache_size_limit();
207 atomic_set(&nn->num_drc_entries, 0);
208 hashsize = nfsd_hashsize(nn->max_drc_entries);
209 nn->maskbits = ilog2(hashsize);
211 nn->nfsd_reply_cache_shrinker.scan_objects = nfsd_reply_cache_scan;
212 nn->nfsd_reply_cache_shrinker.count_objects = nfsd_reply_cache_count;
213 nn->nfsd_reply_cache_shrinker.seeks = 1;
214 status = register_shrinker(&nn->nfsd_reply_cache_shrinker,
215 "nfsd-reply:%s", nn->nfsd_name);
219 nn->drc_hashtbl = kvzalloc(array_size(hashsize,
220 sizeof(*nn->drc_hashtbl)), GFP_KERNEL);
221 if (!nn->drc_hashtbl)
224 for (i = 0; i < hashsize; i++) {
225 INIT_LIST_HEAD(&nn->drc_hashtbl[i].lru_head);
226 spin_lock_init(&nn->drc_hashtbl[i].cache_lock);
228 nn->drc_hashsize = hashsize;
232 unregister_shrinker(&nn->nfsd_reply_cache_shrinker);
233 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
237 void nfsd_reply_cache_shutdown(struct nfsd_net *nn)
239 struct nfsd_cacherep *rp;
242 unregister_shrinker(&nn->nfsd_reply_cache_shrinker);
244 for (i = 0; i < nn->drc_hashsize; i++) {
245 struct list_head *head = &nn->drc_hashtbl[i].lru_head;
246 while (!list_empty(head)) {
247 rp = list_first_entry(head, struct nfsd_cacherep, c_lru);
248 nfsd_reply_cache_free_locked(&nn->drc_hashtbl[i],
253 kvfree(nn->drc_hashtbl);
254 nn->drc_hashtbl = NULL;
255 nn->drc_hashsize = 0;
260 * Move cache entry to end of LRU list, and queue the cleaner to run if it's
261 * not already scheduled.
264 lru_put_end(struct nfsd_drc_bucket *b, struct nfsd_cacherep *rp)
266 rp->c_timestamp = jiffies;
267 list_move_tail(&rp->c_lru, &b->lru_head);
270 static noinline struct nfsd_drc_bucket *
271 nfsd_cache_bucket_find(__be32 xid, struct nfsd_net *nn)
273 unsigned int hash = hash_32((__force u32)xid, nn->maskbits);
275 return &nn->drc_hashtbl[hash];
279 * Remove and return no more than @max expired entries in bucket @b.
280 * If @max is zero, do not limit the number of removed entries.
283 nfsd_prune_bucket_locked(struct nfsd_net *nn, struct nfsd_drc_bucket *b,
284 unsigned int max, struct list_head *dispose)
286 unsigned long expiry = jiffies - RC_EXPIRE;
287 struct nfsd_cacherep *rp, *tmp;
288 unsigned int freed = 0;
290 lockdep_assert_held(&b->cache_lock);
292 /* The bucket LRU is ordered oldest-first. */
293 list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
295 * Don't free entries attached to calls that are still
296 * in-progress, but do keep scanning the list.
298 if (rp->c_state == RC_INPROG)
301 if (atomic_read(&nn->num_drc_entries) <= nn->max_drc_entries &&
302 time_before(expiry, rp->c_timestamp))
305 nfsd_cacherep_unlink_locked(nn, b, rp);
306 list_add(&rp->c_lru, dispose);
308 if (max && ++freed > max)
314 * nfsd_reply_cache_count - count_objects method for the DRC shrinker
315 * @shrink: our registered shrinker context
316 * @sc: garbage collection parameters
318 * Returns the total number of entries in the duplicate reply cache. To
319 * keep things simple and quick, this is not the number of expired entries
320 * in the cache (ie, the number that would be removed by a call to
321 * nfsd_reply_cache_scan).
324 nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
326 struct nfsd_net *nn = container_of(shrink,
327 struct nfsd_net, nfsd_reply_cache_shrinker);
329 return atomic_read(&nn->num_drc_entries);
333 * nfsd_reply_cache_scan - scan_objects method for the DRC shrinker
334 * @shrink: our registered shrinker context
335 * @sc: garbage collection parameters
337 * Free expired entries on each bucket's LRU list until we've released
338 * nr_to_scan freed objects. Nothing will be released if the cache
339 * has not exceeded it's max_drc_entries limit.
341 * Returns the number of entries released by this call.
344 nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
346 struct nfsd_net *nn = container_of(shrink,
347 struct nfsd_net, nfsd_reply_cache_shrinker);
348 unsigned long freed = 0;
352 for (i = 0; i < nn->drc_hashsize; i++) {
353 struct nfsd_drc_bucket *b = &nn->drc_hashtbl[i];
355 if (list_empty(&b->lru_head))
358 spin_lock(&b->cache_lock);
359 nfsd_prune_bucket_locked(nn, b, 0, &dispose);
360 spin_unlock(&b->cache_lock);
362 freed += nfsd_cacherep_dispose(&dispose);
363 if (freed > sc->nr_to_scan)
367 trace_nfsd_drc_gc(nn, freed);
372 * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
375 nfsd_cache_csum(struct svc_rqst *rqstp)
380 struct xdr_buf *buf = &rqstp->rq_arg;
381 const unsigned char *p = buf->head[0].iov_base;
382 size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
384 size_t len = min(buf->head[0].iov_len, csum_len);
386 /* rq_arg.head first */
387 csum = csum_partial(p, len, 0);
390 /* Continue into page array */
391 idx = buf->page_base / PAGE_SIZE;
392 base = buf->page_base & ~PAGE_MASK;
394 p = page_address(buf->pages[idx]) + base;
395 len = min_t(size_t, PAGE_SIZE - base, csum_len);
396 csum = csum_partial(p, len, csum);
405 nfsd_cache_key_cmp(const struct nfsd_cacherep *key,
406 const struct nfsd_cacherep *rp, struct nfsd_net *nn)
408 if (key->c_key.k_xid == rp->c_key.k_xid &&
409 key->c_key.k_csum != rp->c_key.k_csum) {
410 nfsd_stats_payload_misses_inc(nn);
411 trace_nfsd_drc_mismatch(nn, key, rp);
414 return memcmp(&key->c_key, &rp->c_key, sizeof(key->c_key));
418 * Search the request hash for an entry that matches the given rqstp.
419 * Must be called with cache_lock held. Returns the found entry or
420 * inserts an empty key on failure.
422 static struct nfsd_cacherep *
423 nfsd_cache_insert(struct nfsd_drc_bucket *b, struct nfsd_cacherep *key,
426 struct nfsd_cacherep *rp, *ret = key;
427 struct rb_node **p = &b->rb_head.rb_node,
429 unsigned int entries = 0;
435 rp = rb_entry(parent, struct nfsd_cacherep, c_node);
437 cmp = nfsd_cache_key_cmp(key, rp, nn);
439 p = &parent->rb_left;
441 p = &parent->rb_right;
447 rb_link_node(&key->c_node, parent, p);
448 rb_insert_color(&key->c_node, &b->rb_head);
450 /* tally hash chain length stats */
451 if (entries > nn->longest_chain) {
452 nn->longest_chain = entries;
453 nn->longest_chain_cachesize = atomic_read(&nn->num_drc_entries);
454 } else if (entries == nn->longest_chain) {
455 /* prefer to keep the smallest cachesize possible here */
456 nn->longest_chain_cachesize = min_t(unsigned int,
457 nn->longest_chain_cachesize,
458 atomic_read(&nn->num_drc_entries));
466 * nfsd_cache_lookup - Find an entry in the duplicate reply cache
467 * @rqstp: Incoming Call to find
468 * @cacherep: OUT: DRC entry for this request
470 * Try to find an entry matching the current call in the cache. When none
471 * is found, we try to grab the oldest expired entry off the LRU list. If
472 * a suitable one isn't there, then drop the cache_lock and allocate a
473 * new one, then search again in case one got inserted while this thread
474 * didn't hold the lock.
477 * %RC_DOIT: Process the request normally
478 * %RC_REPLY: Reply from cache
479 * %RC_DROPIT: Do not process the request further
481 int nfsd_cache_lookup(struct svc_rqst *rqstp, struct nfsd_cacherep **cacherep)
484 struct nfsd_cacherep *rp, *found;
486 struct nfsd_drc_bucket *b;
487 int type = rqstp->rq_cachetype;
492 if (type == RC_NOCACHE) {
493 nfsd_stats_rc_nocache_inc();
497 csum = nfsd_cache_csum(rqstp);
500 * Since the common case is a cache miss followed by an insert,
501 * preallocate an entry.
503 nn = net_generic(SVC_NET(rqstp), nfsd_net_id);
504 rp = nfsd_cacherep_alloc(rqstp, csum, nn);
508 b = nfsd_cache_bucket_find(rqstp->rq_xid, nn);
509 spin_lock(&b->cache_lock);
510 found = nfsd_cache_insert(b, rp, nn);
514 rp->c_state = RC_INPROG;
515 nfsd_prune_bucket_locked(nn, b, 3, &dispose);
516 spin_unlock(&b->cache_lock);
518 freed = nfsd_cacherep_dispose(&dispose);
519 trace_nfsd_drc_gc(nn, freed);
521 nfsd_stats_rc_misses_inc();
522 atomic_inc(&nn->num_drc_entries);
523 nfsd_stats_drc_mem_usage_add(nn, sizeof(*rp));
527 /* We found a matching entry which is either in progress or done. */
528 nfsd_reply_cache_free_locked(NULL, rp, nn);
529 nfsd_stats_rc_hits_inc();
533 /* Request being processed */
534 if (rp->c_state == RC_INPROG)
537 /* From the hall of fame of impractical attacks:
538 * Is this a user who tries to snoop on the cache? */
540 if (!test_bit(RQ_SECURE, &rqstp->rq_flags) && rp->c_secure)
543 /* Compose RPC reply header */
544 switch (rp->c_type) {
548 xdr_stream_encode_be32(&rqstp->rq_res_stream, rp->c_replstat);
552 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
553 goto out_unlock; /* should not happen */
557 WARN_ONCE(1, "nfsd: bad repcache type %d\n", rp->c_type);
561 trace_nfsd_drc_found(nn, rqstp, rtn);
563 spin_unlock(&b->cache_lock);
569 * nfsd_cache_update - Update an entry in the duplicate reply cache.
570 * @rqstp: svc_rqst with a finished Reply
571 * @rp: IN: DRC entry for this request
572 * @cachetype: which cache to update
573 * @statp: pointer to Reply's NFS status code, or NULL
575 * This is called from nfsd_dispatch when the procedure has been
576 * executed and the complete reply is in rqstp->rq_res.
578 * We're copying around data here rather than swapping buffers because
579 * the toplevel loop requires max-sized buffers, which would be a waste
580 * of memory for a cache with a max reply size of 100 bytes (diropokres).
582 * If we should start to use different types of cache entries tailored
583 * specifically for attrstat and fh's, we may save even more space.
585 * Also note that a cachetype of RC_NOCACHE can legally be passed when
586 * nfsd failed to encode a reply that otherwise would have been cached.
587 * In this case, nfsd_cache_update is called with statp == NULL.
589 void nfsd_cache_update(struct svc_rqst *rqstp, struct nfsd_cacherep *rp,
590 int cachetype, __be32 *statp)
592 struct nfsd_net *nn = net_generic(SVC_NET(rqstp), nfsd_net_id);
593 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
594 struct nfsd_drc_bucket *b;
601 b = nfsd_cache_bucket_find(rp->c_key.k_xid, nn);
603 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
606 /* Don't cache excessive amounts of data and XDR failures */
607 if (!statp || len > (256 >> 2)) {
608 nfsd_reply_cache_free(b, rp, nn);
615 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
616 rp->c_replstat = *statp;
619 cachv = &rp->c_replvec;
621 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
622 if (!cachv->iov_base) {
623 nfsd_reply_cache_free(b, rp, nn);
626 cachv->iov_len = bufsize;
627 memcpy(cachv->iov_base, statp, bufsize);
630 nfsd_reply_cache_free(b, rp, nn);
633 spin_lock(&b->cache_lock);
634 nfsd_stats_drc_mem_usage_add(nn, bufsize);
636 rp->c_secure = test_bit(RQ_SECURE, &rqstp->rq_flags);
637 rp->c_type = cachetype;
638 rp->c_state = RC_DONE;
639 spin_unlock(&b->cache_lock);
644 * Copy cached reply to current reply buffer. Should always fit.
645 * FIXME as reply is in a page, we should just attach the page, and
646 * keep a refcount....
649 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
651 struct kvec *vec = &rqstp->rq_res.head[0];
653 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
654 printk(KERN_WARNING "nfsd: cached reply too large (%zd).\n",
658 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
659 vec->iov_len += data->iov_len;
664 * Note that fields may be added, removed or reordered in the future. Programs
665 * scraping this file for info should test the labels to ensure they're
666 * getting the correct field.
668 int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
670 struct nfsd_net *nn = net_generic(file_inode(m->file)->i_sb->s_fs_info,
673 seq_printf(m, "max entries: %u\n", nn->max_drc_entries);
674 seq_printf(m, "num entries: %u\n",
675 atomic_read(&nn->num_drc_entries));
676 seq_printf(m, "hash buckets: %u\n", 1 << nn->maskbits);
677 seq_printf(m, "mem usage: %lld\n",
678 percpu_counter_sum_positive(&nn->counter[NFSD_NET_DRC_MEM_USAGE]));
679 seq_printf(m, "cache hits: %lld\n",
680 percpu_counter_sum_positive(&nfsdstats.counter[NFSD_STATS_RC_HITS]));
681 seq_printf(m, "cache misses: %lld\n",
682 percpu_counter_sum_positive(&nfsdstats.counter[NFSD_STATS_RC_MISSES]));
683 seq_printf(m, "not cached: %lld\n",
684 percpu_counter_sum_positive(&nfsdstats.counter[NFSD_STATS_RC_NOCACHE]));
685 seq_printf(m, "payload misses: %lld\n",
686 percpu_counter_sum_positive(&nn->counter[NFSD_NET_PAYLOAD_MISSES]));
687 seq_printf(m, "longest chain len: %u\n", nn->longest_chain);
688 seq_printf(m, "cachesize at longest: %u\n", nn->longest_chain_cachesize);