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 svc_cacherep *
88 nfsd_reply_cache_alloc(struct svc_rqst *rqstp, __wsum csum,
91 struct svc_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;
114 nfsd_reply_cache_free_locked(struct nfsd_drc_bucket *b, struct svc_cacherep *rp,
117 if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
118 nfsd_stats_drc_mem_usage_sub(nn, rp->c_replvec.iov_len);
119 kfree(rp->c_replvec.iov_base);
121 if (rp->c_state != RC_UNUSED) {
122 rb_erase(&rp->c_node, &b->rb_head);
123 list_del(&rp->c_lru);
124 atomic_dec(&nn->num_drc_entries);
125 nfsd_stats_drc_mem_usage_sub(nn, sizeof(*rp));
127 kmem_cache_free(drc_slab, rp);
131 nfsd_reply_cache_free(struct nfsd_drc_bucket *b, struct svc_cacherep *rp,
134 spin_lock(&b->cache_lock);
135 nfsd_reply_cache_free_locked(b, rp, nn);
136 spin_unlock(&b->cache_lock);
139 int nfsd_drc_slab_create(void)
141 drc_slab = kmem_cache_create("nfsd_drc",
142 sizeof(struct svc_cacherep), 0, 0, NULL);
143 return drc_slab ? 0: -ENOMEM;
146 void nfsd_drc_slab_free(void)
148 kmem_cache_destroy(drc_slab);
151 static int nfsd_reply_cache_stats_init(struct nfsd_net *nn)
153 return nfsd_percpu_counters_init(nn->counter, NFSD_NET_COUNTERS_NUM);
156 static void nfsd_reply_cache_stats_destroy(struct nfsd_net *nn)
158 nfsd_percpu_counters_destroy(nn->counter, NFSD_NET_COUNTERS_NUM);
161 int nfsd_reply_cache_init(struct nfsd_net *nn)
163 unsigned int hashsize;
167 nn->max_drc_entries = nfsd_cache_size_limit();
168 atomic_set(&nn->num_drc_entries, 0);
169 hashsize = nfsd_hashsize(nn->max_drc_entries);
170 nn->maskbits = ilog2(hashsize);
172 status = nfsd_reply_cache_stats_init(nn);
176 nn->nfsd_reply_cache_shrinker.scan_objects = nfsd_reply_cache_scan;
177 nn->nfsd_reply_cache_shrinker.count_objects = nfsd_reply_cache_count;
178 nn->nfsd_reply_cache_shrinker.seeks = 1;
179 status = register_shrinker(&nn->nfsd_reply_cache_shrinker,
180 "nfsd-reply:%s", nn->nfsd_name);
182 goto out_stats_destroy;
184 nn->drc_hashtbl = kvzalloc(array_size(hashsize,
185 sizeof(*nn->drc_hashtbl)), GFP_KERNEL);
186 if (!nn->drc_hashtbl)
189 for (i = 0; i < hashsize; i++) {
190 INIT_LIST_HEAD(&nn->drc_hashtbl[i].lru_head);
191 spin_lock_init(&nn->drc_hashtbl[i].cache_lock);
193 nn->drc_hashsize = hashsize;
197 unregister_shrinker(&nn->nfsd_reply_cache_shrinker);
199 nfsd_reply_cache_stats_destroy(nn);
201 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
205 void nfsd_reply_cache_shutdown(struct nfsd_net *nn)
207 struct svc_cacherep *rp;
210 unregister_shrinker(&nn->nfsd_reply_cache_shrinker);
212 for (i = 0; i < nn->drc_hashsize; i++) {
213 struct list_head *head = &nn->drc_hashtbl[i].lru_head;
214 while (!list_empty(head)) {
215 rp = list_first_entry(head, struct svc_cacherep, c_lru);
216 nfsd_reply_cache_free_locked(&nn->drc_hashtbl[i],
220 nfsd_reply_cache_stats_destroy(nn);
222 kvfree(nn->drc_hashtbl);
223 nn->drc_hashtbl = NULL;
224 nn->drc_hashsize = 0;
229 * Move cache entry to end of LRU list, and queue the cleaner to run if it's
230 * not already scheduled.
233 lru_put_end(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
235 rp->c_timestamp = jiffies;
236 list_move_tail(&rp->c_lru, &b->lru_head);
239 static noinline struct nfsd_drc_bucket *
240 nfsd_cache_bucket_find(__be32 xid, struct nfsd_net *nn)
242 unsigned int hash = hash_32((__force u32)xid, nn->maskbits);
244 return &nn->drc_hashtbl[hash];
247 static long prune_bucket(struct nfsd_drc_bucket *b, struct nfsd_net *nn,
250 struct svc_cacherep *rp, *tmp;
253 list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
255 * Don't free entries attached to calls that are still
256 * in-progress, but do keep scanning the list.
258 if (rp->c_state == RC_INPROG)
260 if (atomic_read(&nn->num_drc_entries) <= nn->max_drc_entries &&
261 time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
263 nfsd_reply_cache_free_locked(b, rp, nn);
264 if (max && freed++ > max)
270 static long nfsd_prune_bucket(struct nfsd_drc_bucket *b, struct nfsd_net *nn)
272 return prune_bucket(b, nn, 3);
276 * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
277 * Also prune the oldest ones when the total exceeds the max number of entries.
280 prune_cache_entries(struct nfsd_net *nn)
285 for (i = 0; i < nn->drc_hashsize; i++) {
286 struct nfsd_drc_bucket *b = &nn->drc_hashtbl[i];
288 if (list_empty(&b->lru_head))
290 spin_lock(&b->cache_lock);
291 freed += prune_bucket(b, nn, 0);
292 spin_unlock(&b->cache_lock);
298 nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
300 struct nfsd_net *nn = container_of(shrink,
301 struct nfsd_net, nfsd_reply_cache_shrinker);
303 return atomic_read(&nn->num_drc_entries);
307 nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
309 struct nfsd_net *nn = container_of(shrink,
310 struct nfsd_net, nfsd_reply_cache_shrinker);
312 return prune_cache_entries(nn);
316 * nfsd_cache_csum - Checksum incoming NFS Call arguments
317 * @buf: buffer containing a whole RPC Call message
318 * @start: starting byte of the NFS Call header
319 * @remaining: size of the NFS Call header, in bytes
321 * Compute a weak checksum of the leading bytes of an NFS procedure
322 * call header to help verify that a retransmitted Call matches an
323 * entry in the duplicate reply cache.
325 * To avoid assumptions about how the RPC message is laid out in
326 * @buf and what else it might contain (eg, a GSS MIC suffix), the
327 * caller passes us the exact location and length of the NFS Call
330 * Returns a 32-bit checksum value, as defined in RFC 793.
332 static __wsum nfsd_cache_csum(struct xdr_buf *buf, unsigned int start,
333 unsigned int remaining)
335 unsigned int base, len;
336 struct xdr_buf subbuf;
341 if (remaining > RC_CSUMLEN)
342 remaining = RC_CSUMLEN;
343 if (xdr_buf_subsegment(buf, &subbuf, start, remaining))
346 /* rq_arg.head first */
347 if (subbuf.head[0].iov_len) {
348 len = min_t(unsigned int, subbuf.head[0].iov_len, remaining);
349 csum = csum_partial(subbuf.head[0].iov_base, len, csum);
353 /* Continue into page array */
354 idx = subbuf.page_base / PAGE_SIZE;
355 base = subbuf.page_base & ~PAGE_MASK;
357 p = page_address(subbuf.pages[idx]) + base;
358 len = min_t(unsigned int, PAGE_SIZE - base, remaining);
359 csum = csum_partial(p, len, csum);
368 nfsd_cache_key_cmp(const struct svc_cacherep *key,
369 const struct svc_cacherep *rp, struct nfsd_net *nn)
371 if (key->c_key.k_xid == rp->c_key.k_xid &&
372 key->c_key.k_csum != rp->c_key.k_csum) {
373 nfsd_stats_payload_misses_inc(nn);
374 trace_nfsd_drc_mismatch(nn, key, rp);
377 return memcmp(&key->c_key, &rp->c_key, sizeof(key->c_key));
381 * Search the request hash for an entry that matches the given rqstp.
382 * Must be called with cache_lock held. Returns the found entry or
383 * inserts an empty key on failure.
385 static struct svc_cacherep *
386 nfsd_cache_insert(struct nfsd_drc_bucket *b, struct svc_cacherep *key,
389 struct svc_cacherep *rp, *ret = key;
390 struct rb_node **p = &b->rb_head.rb_node,
392 unsigned int entries = 0;
398 rp = rb_entry(parent, struct svc_cacherep, c_node);
400 cmp = nfsd_cache_key_cmp(key, rp, nn);
402 p = &parent->rb_left;
404 p = &parent->rb_right;
410 rb_link_node(&key->c_node, parent, p);
411 rb_insert_color(&key->c_node, &b->rb_head);
413 /* tally hash chain length stats */
414 if (entries > nn->longest_chain) {
415 nn->longest_chain = entries;
416 nn->longest_chain_cachesize = atomic_read(&nn->num_drc_entries);
417 } else if (entries == nn->longest_chain) {
418 /* prefer to keep the smallest cachesize possible here */
419 nn->longest_chain_cachesize = min_t(unsigned int,
420 nn->longest_chain_cachesize,
421 atomic_read(&nn->num_drc_entries));
429 * nfsd_cache_lookup - Find an entry in the duplicate reply cache
430 * @rqstp: Incoming Call to find
431 * @start: starting byte in @rqstp->rq_arg of the NFS Call header
432 * @len: size of the NFS Call header, in bytes
434 * Try to find an entry matching the current call in the cache. When none
435 * is found, we try to grab the oldest expired entry off the LRU list. If
436 * a suitable one isn't there, then drop the cache_lock and allocate a
437 * new one, then search again in case one got inserted while this thread
438 * didn't hold the lock.
441 * %RC_DOIT: Process the request normally
442 * %RC_REPLY: Reply from cache
443 * %RC_DROPIT: Do not process the request further
445 int nfsd_cache_lookup(struct svc_rqst *rqstp, unsigned int start,
449 struct svc_cacherep *rp, *found;
451 struct nfsd_drc_bucket *b;
452 int type = rqstp->rq_cachetype;
455 rqstp->rq_cacherep = NULL;
456 if (type == RC_NOCACHE) {
457 nfsd_stats_rc_nocache_inc();
461 csum = nfsd_cache_csum(&rqstp->rq_arg, start, len);
464 * Since the common case is a cache miss followed by an insert,
465 * preallocate an entry.
467 nn = net_generic(SVC_NET(rqstp), nfsd_net_id);
468 rp = nfsd_reply_cache_alloc(rqstp, csum, nn);
472 b = nfsd_cache_bucket_find(rqstp->rq_xid, nn);
473 spin_lock(&b->cache_lock);
474 found = nfsd_cache_insert(b, rp, nn);
478 nfsd_stats_rc_misses_inc();
479 rqstp->rq_cacherep = rp;
480 rp->c_state = RC_INPROG;
482 atomic_inc(&nn->num_drc_entries);
483 nfsd_stats_drc_mem_usage_add(nn, sizeof(*rp));
485 nfsd_prune_bucket(b, nn);
488 spin_unlock(&b->cache_lock);
493 /* We found a matching entry which is either in progress or done. */
494 nfsd_reply_cache_free_locked(NULL, rp, nn);
495 nfsd_stats_rc_hits_inc();
499 /* Request being processed */
500 if (rp->c_state == RC_INPROG)
503 /* From the hall of fame of impractical attacks:
504 * Is this a user who tries to snoop on the cache? */
506 if (!test_bit(RQ_SECURE, &rqstp->rq_flags) && rp->c_secure)
509 /* Compose RPC reply header */
510 switch (rp->c_type) {
514 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
518 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
519 goto out_unlock; /* should not happen */
523 WARN_ONCE(1, "nfsd: bad repcache type %d\n", rp->c_type);
527 trace_nfsd_drc_found(nn, rqstp, rtn);
532 * nfsd_cache_update - Update an entry in the duplicate reply cache.
533 * @rqstp: svc_rqst with a finished Reply
534 * @cachetype: which cache to update
535 * @statp: Reply's status code
537 * This is called from nfsd_dispatch when the procedure has been
538 * executed and the complete reply is in rqstp->rq_res.
540 * We're copying around data here rather than swapping buffers because
541 * the toplevel loop requires max-sized buffers, which would be a waste
542 * of memory for a cache with a max reply size of 100 bytes (diropokres).
544 * If we should start to use different types of cache entries tailored
545 * specifically for attrstat and fh's, we may save even more space.
547 * Also note that a cachetype of RC_NOCACHE can legally be passed when
548 * nfsd failed to encode a reply that otherwise would have been cached.
549 * In this case, nfsd_cache_update is called with statp == NULL.
551 void nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
553 struct nfsd_net *nn = net_generic(SVC_NET(rqstp), nfsd_net_id);
554 struct svc_cacherep *rp = rqstp->rq_cacherep;
555 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
556 struct nfsd_drc_bucket *b;
563 b = nfsd_cache_bucket_find(rp->c_key.k_xid, nn);
565 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
568 /* Don't cache excessive amounts of data and XDR failures */
569 if (!statp || len > (256 >> 2)) {
570 nfsd_reply_cache_free(b, rp, nn);
577 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
578 rp->c_replstat = *statp;
581 cachv = &rp->c_replvec;
583 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
584 if (!cachv->iov_base) {
585 nfsd_reply_cache_free(b, rp, nn);
588 cachv->iov_len = bufsize;
589 memcpy(cachv->iov_base, statp, bufsize);
592 nfsd_reply_cache_free(b, rp, nn);
595 spin_lock(&b->cache_lock);
596 nfsd_stats_drc_mem_usage_add(nn, bufsize);
598 rp->c_secure = test_bit(RQ_SECURE, &rqstp->rq_flags);
599 rp->c_type = cachetype;
600 rp->c_state = RC_DONE;
601 spin_unlock(&b->cache_lock);
606 * Copy cached reply to current reply buffer. Should always fit.
607 * FIXME as reply is in a page, we should just attach the page, and
608 * keep a refcount....
611 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
613 struct kvec *vec = &rqstp->rq_res.head[0];
615 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
616 printk(KERN_WARNING "nfsd: cached reply too large (%zd).\n",
620 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
621 vec->iov_len += data->iov_len;
626 * Note that fields may be added, removed or reordered in the future. Programs
627 * scraping this file for info should test the labels to ensure they're
628 * getting the correct field.
630 int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
632 struct nfsd_net *nn = net_generic(file_inode(m->file)->i_sb->s_fs_info,
635 seq_printf(m, "max entries: %u\n", nn->max_drc_entries);
636 seq_printf(m, "num entries: %u\n",
637 atomic_read(&nn->num_drc_entries));
638 seq_printf(m, "hash buckets: %u\n", 1 << nn->maskbits);
639 seq_printf(m, "mem usage: %lld\n",
640 percpu_counter_sum_positive(&nn->counter[NFSD_NET_DRC_MEM_USAGE]));
641 seq_printf(m, "cache hits: %lld\n",
642 percpu_counter_sum_positive(&nfsdstats.counter[NFSD_STATS_RC_HITS]));
643 seq_printf(m, "cache misses: %lld\n",
644 percpu_counter_sum_positive(&nfsdstats.counter[NFSD_STATS_RC_MISSES]));
645 seq_printf(m, "not cached: %lld\n",
646 percpu_counter_sum_positive(&nfsdstats.counter[NFSD_STATS_RC_NOCACHE]));
647 seq_printf(m, "payload misses: %lld\n",
648 percpu_counter_sum_positive(&nn->counter[NFSD_NET_PAYLOAD_MISSES]));
649 seq_printf(m, "longest chain len: %u\n", nn->longest_chain);
650 seq_printf(m, "cachesize at longest: %u\n", nn->longest_chain_cachesize);