Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Request reply cache. This is currently a global cache, but this may
4 * change in the future and be a per-client cache.
5 *
6 * This code is heavily inspired by the 44BSD implementation, although
7 * it does things a bit differently.
8 *
9 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
10 */
11
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>
20
21#include "nfsd.h"
22#include "cache.h"
23#include "trace.h"
24
25/*
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.
29 */
30#define TARGET_BUCKET_SIZE 64
31
32struct nfsd_drc_bucket {
33 struct rb_root rb_head;
34 struct list_head lru_head;
35 spinlock_t cache_lock;
36};
37
38static struct kmem_cache *drc_slab;
39
40static int nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
41static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
42 struct shrink_control *sc);
43static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
44 struct shrink_control *sc);
45
46/*
47 * Put a cap on the size of the DRC based on the amount of available
48 * low memory in the machine.
49 *
50 * 64MB: 8192
51 * 128MB: 11585
52 * 256MB: 16384
53 * 512MB: 23170
54 * 1GB: 32768
55 * 2GB: 46340
56 * 4GB: 65536
57 * 8GB: 92681
58 * 16GB: 131072
59 *
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
62 * used in k.
63 *
64 * XXX: these limits are per-container, so memory used will increase
65 * linearly with number of containers. Maybe that's OK.
66 */
67static unsigned int
68nfsd_cache_size_limit(void)
69{
70 unsigned int limit;
71 unsigned long low_pages = totalram_pages() - totalhigh_pages();
72
73 limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
74 return min_t(unsigned int, limit, 256*1024);
75}
76
77/*
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.
80 */
81static unsigned int
82nfsd_hashsize(unsigned int limit)
83{
84 return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
85}
86
87static struct nfsd_cacherep *
88nfsd_cacherep_alloc(struct svc_rqst *rqstp, __wsum csum,
89 struct nfsd_net *nn)
90{
91 struct nfsd_cacherep *rp;
92
93 rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
94 if (rp) {
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);
99
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;
109 }
110 return rp;
111}
112
113static void nfsd_cacherep_free(struct nfsd_cacherep *rp)
114{
115 if (rp->c_type == RC_REPLBUFF)
116 kfree(rp->c_replvec.iov_base);
117 kmem_cache_free(drc_slab, rp);
118}
119
120static unsigned long
121nfsd_cacherep_dispose(struct list_head *dispose)
122{
123 struct nfsd_cacherep *rp;
124 unsigned long freed = 0;
125
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);
130 freed++;
131 }
132 return freed;
133}
134
135static void
136nfsd_cacherep_unlink_locked(struct nfsd_net *nn, struct nfsd_drc_bucket *b,
137 struct nfsd_cacherep *rp)
138{
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));
146 }
147}
148
149static void
150nfsd_reply_cache_free_locked(struct nfsd_drc_bucket *b, struct nfsd_cacherep *rp,
151 struct nfsd_net *nn)
152{
153 nfsd_cacherep_unlink_locked(nn, b, rp);
154 nfsd_cacherep_free(rp);
155}
156
157static void
158nfsd_reply_cache_free(struct nfsd_drc_bucket *b, struct nfsd_cacherep *rp,
159 struct nfsd_net *nn)
160{
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);
165}
166
167int nfsd_drc_slab_create(void)
168{
169 drc_slab = KMEM_CACHE(nfsd_cacherep, 0);
170 return drc_slab ? 0: -ENOMEM;
171}
172
173void nfsd_drc_slab_free(void)
174{
175 kmem_cache_destroy(drc_slab);
176}
177
178int nfsd_reply_cache_init(struct nfsd_net *nn)
179{
180 unsigned int hashsize;
181 unsigned int i;
182
183 nn->max_drc_entries = nfsd_cache_size_limit();
184 atomic_set(&nn->num_drc_entries, 0);
185 hashsize = nfsd_hashsize(nn->max_drc_entries);
186 nn->maskbits = ilog2(hashsize);
187
188 nn->drc_hashtbl = kvzalloc(array_size(hashsize,
189 sizeof(*nn->drc_hashtbl)), GFP_KERNEL);
190 if (!nn->drc_hashtbl)
191 return -ENOMEM;
192
193 nn->nfsd_reply_cache_shrinker = shrinker_alloc(0, "nfsd-reply:%s",
194 nn->nfsd_name);
195 if (!nn->nfsd_reply_cache_shrinker)
196 goto out_shrinker;
197
198 nn->nfsd_reply_cache_shrinker->scan_objects = nfsd_reply_cache_scan;
199 nn->nfsd_reply_cache_shrinker->count_objects = nfsd_reply_cache_count;
200 nn->nfsd_reply_cache_shrinker->seeks = 1;
201 nn->nfsd_reply_cache_shrinker->private_data = nn;
202
203 shrinker_register(nn->nfsd_reply_cache_shrinker);
204
205 for (i = 0; i < hashsize; i++) {
206 INIT_LIST_HEAD(&nn->drc_hashtbl[i].lru_head);
207 spin_lock_init(&nn->drc_hashtbl[i].cache_lock);
208 }
209 nn->drc_hashsize = hashsize;
210
211 return 0;
212out_shrinker:
213 kvfree(nn->drc_hashtbl);
214 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
215 return -ENOMEM;
216}
217
218void nfsd_reply_cache_shutdown(struct nfsd_net *nn)
219{
220 struct nfsd_cacherep *rp;
221 unsigned int i;
222
223 shrinker_free(nn->nfsd_reply_cache_shrinker);
224
225 for (i = 0; i < nn->drc_hashsize; i++) {
226 struct list_head *head = &nn->drc_hashtbl[i].lru_head;
227 while (!list_empty(head)) {
228 rp = list_first_entry(head, struct nfsd_cacherep, c_lru);
229 nfsd_reply_cache_free_locked(&nn->drc_hashtbl[i],
230 rp, nn);
231 }
232 }
233
234 kvfree(nn->drc_hashtbl);
235 nn->drc_hashtbl = NULL;
236 nn->drc_hashsize = 0;
237
238}
239
240/*
241 * Move cache entry to end of LRU list, and queue the cleaner to run if it's
242 * not already scheduled.
243 */
244static void
245lru_put_end(struct nfsd_drc_bucket *b, struct nfsd_cacherep *rp)
246{
247 rp->c_timestamp = jiffies;
248 list_move_tail(&rp->c_lru, &b->lru_head);
249}
250
251static noinline struct nfsd_drc_bucket *
252nfsd_cache_bucket_find(__be32 xid, struct nfsd_net *nn)
253{
254 unsigned int hash = hash_32((__force u32)xid, nn->maskbits);
255
256 return &nn->drc_hashtbl[hash];
257}
258
259/*
260 * Remove and return no more than @max expired entries in bucket @b.
261 * If @max is zero, do not limit the number of removed entries.
262 */
263static void
264nfsd_prune_bucket_locked(struct nfsd_net *nn, struct nfsd_drc_bucket *b,
265 unsigned int max, struct list_head *dispose)
266{
267 unsigned long expiry = jiffies - RC_EXPIRE;
268 struct nfsd_cacherep *rp, *tmp;
269 unsigned int freed = 0;
270
271 lockdep_assert_held(&b->cache_lock);
272
273 /* The bucket LRU is ordered oldest-first. */
274 list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
275 /*
276 * Don't free entries attached to calls that are still
277 * in-progress, but do keep scanning the list.
278 */
279 if (rp->c_state == RC_INPROG)
280 continue;
281
282 if (atomic_read(&nn->num_drc_entries) <= nn->max_drc_entries &&
283 time_before(expiry, rp->c_timestamp))
284 break;
285
286 nfsd_cacherep_unlink_locked(nn, b, rp);
287 list_add(&rp->c_lru, dispose);
288
289 if (max && ++freed > max)
290 break;
291 }
292}
293
294/**
295 * nfsd_reply_cache_count - count_objects method for the DRC shrinker
296 * @shrink: our registered shrinker context
297 * @sc: garbage collection parameters
298 *
299 * Returns the total number of entries in the duplicate reply cache. To
300 * keep things simple and quick, this is not the number of expired entries
301 * in the cache (ie, the number that would be removed by a call to
302 * nfsd_reply_cache_scan).
303 */
304static unsigned long
305nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
306{
307 struct nfsd_net *nn = shrink->private_data;
308
309 return atomic_read(&nn->num_drc_entries);
310}
311
312/**
313 * nfsd_reply_cache_scan - scan_objects method for the DRC shrinker
314 * @shrink: our registered shrinker context
315 * @sc: garbage collection parameters
316 *
317 * Free expired entries on each bucket's LRU list until we've released
318 * nr_to_scan freed objects. Nothing will be released if the cache
319 * has not exceeded it's max_drc_entries limit.
320 *
321 * Returns the number of entries released by this call.
322 */
323static unsigned long
324nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
325{
326 struct nfsd_net *nn = shrink->private_data;
327 unsigned long freed = 0;
328 LIST_HEAD(dispose);
329 unsigned int i;
330
331 for (i = 0; i < nn->drc_hashsize; i++) {
332 struct nfsd_drc_bucket *b = &nn->drc_hashtbl[i];
333
334 if (list_empty(&b->lru_head))
335 continue;
336
337 spin_lock(&b->cache_lock);
338 nfsd_prune_bucket_locked(nn, b, 0, &dispose);
339 spin_unlock(&b->cache_lock);
340
341 freed += nfsd_cacherep_dispose(&dispose);
342 if (freed > sc->nr_to_scan)
343 break;
344 }
345 return freed;
346}
347
348/**
349 * nfsd_cache_csum - Checksum incoming NFS Call arguments
350 * @buf: buffer containing a whole RPC Call message
351 * @start: starting byte of the NFS Call header
352 * @remaining: size of the NFS Call header, in bytes
353 *
354 * Compute a weak checksum of the leading bytes of an NFS procedure
355 * call header to help verify that a retransmitted Call matches an
356 * entry in the duplicate reply cache.
357 *
358 * To avoid assumptions about how the RPC message is laid out in
359 * @buf and what else it might contain (eg, a GSS MIC suffix), the
360 * caller passes us the exact location and length of the NFS Call
361 * header.
362 *
363 * Returns a 32-bit checksum value, as defined in RFC 793.
364 */
365static __wsum nfsd_cache_csum(struct xdr_buf *buf, unsigned int start,
366 unsigned int remaining)
367{
368 unsigned int base, len;
369 struct xdr_buf subbuf;
370 __wsum csum = 0;
371 void *p;
372 int idx;
373
374 if (remaining > RC_CSUMLEN)
375 remaining = RC_CSUMLEN;
376 if (xdr_buf_subsegment(buf, &subbuf, start, remaining))
377 return csum;
378
379 /* rq_arg.head first */
380 if (subbuf.head[0].iov_len) {
381 len = min_t(unsigned int, subbuf.head[0].iov_len, remaining);
382 csum = csum_partial(subbuf.head[0].iov_base, len, csum);
383 remaining -= len;
384 }
385
386 /* Continue into page array */
387 idx = subbuf.page_base / PAGE_SIZE;
388 base = subbuf.page_base & ~PAGE_MASK;
389 while (remaining) {
390 p = page_address(subbuf.pages[idx]) + base;
391 len = min_t(unsigned int, PAGE_SIZE - base, remaining);
392 csum = csum_partial(p, len, csum);
393 remaining -= len;
394 base = 0;
395 ++idx;
396 }
397 return csum;
398}
399
400static int
401nfsd_cache_key_cmp(const struct nfsd_cacherep *key,
402 const struct nfsd_cacherep *rp, struct nfsd_net *nn)
403{
404 if (key->c_key.k_xid == rp->c_key.k_xid &&
405 key->c_key.k_csum != rp->c_key.k_csum) {
406 nfsd_stats_payload_misses_inc(nn);
407 trace_nfsd_drc_mismatch(nn, key, rp);
408 }
409
410 return memcmp(&key->c_key, &rp->c_key, sizeof(key->c_key));
411}
412
413/*
414 * Search the request hash for an entry that matches the given rqstp.
415 * Must be called with cache_lock held. Returns the found entry or
416 * inserts an empty key on failure.
417 */
418static struct nfsd_cacherep *
419nfsd_cache_insert(struct nfsd_drc_bucket *b, struct nfsd_cacherep *key,
420 struct nfsd_net *nn)
421{
422 struct nfsd_cacherep *rp, *ret = key;
423 struct rb_node **p = &b->rb_head.rb_node,
424 *parent = NULL;
425 unsigned int entries = 0;
426 int cmp;
427
428 while (*p != NULL) {
429 ++entries;
430 parent = *p;
431 rp = rb_entry(parent, struct nfsd_cacherep, c_node);
432
433 cmp = nfsd_cache_key_cmp(key, rp, nn);
434 if (cmp < 0)
435 p = &parent->rb_left;
436 else if (cmp > 0)
437 p = &parent->rb_right;
438 else {
439 ret = rp;
440 goto out;
441 }
442 }
443 rb_link_node(&key->c_node, parent, p);
444 rb_insert_color(&key->c_node, &b->rb_head);
445out:
446 /* tally hash chain length stats */
447 if (entries > nn->longest_chain) {
448 nn->longest_chain = entries;
449 nn->longest_chain_cachesize = atomic_read(&nn->num_drc_entries);
450 } else if (entries == nn->longest_chain) {
451 /* prefer to keep the smallest cachesize possible here */
452 nn->longest_chain_cachesize = min_t(unsigned int,
453 nn->longest_chain_cachesize,
454 atomic_read(&nn->num_drc_entries));
455 }
456
457 lru_put_end(b, ret);
458 return ret;
459}
460
461/**
462 * nfsd_cache_lookup - Find an entry in the duplicate reply cache
463 * @rqstp: Incoming Call to find
464 * @start: starting byte in @rqstp->rq_arg of the NFS Call header
465 * @len: size of the NFS Call header, in bytes
466 * @cacherep: OUT: DRC entry for this request
467 *
468 * Try to find an entry matching the current call in the cache. When none
469 * is found, we try to grab the oldest expired entry off the LRU list. If
470 * a suitable one isn't there, then drop the cache_lock and allocate a
471 * new one, then search again in case one got inserted while this thread
472 * didn't hold the lock.
473 *
474 * Return values:
475 * %RC_DOIT: Process the request normally
476 * %RC_REPLY: Reply from cache
477 * %RC_DROPIT: Do not process the request further
478 */
479int nfsd_cache_lookup(struct svc_rqst *rqstp, unsigned int start,
480 unsigned int len, struct nfsd_cacherep **cacherep)
481{
482 struct nfsd_net *nn = net_generic(SVC_NET(rqstp), nfsd_net_id);
483 struct nfsd_cacherep *rp, *found;
484 __wsum csum;
485 struct nfsd_drc_bucket *b;
486 int type = rqstp->rq_cachetype;
487 LIST_HEAD(dispose);
488 int rtn = RC_DOIT;
489
490 if (type == RC_NOCACHE) {
491 nfsd_stats_rc_nocache_inc(nn);
492 goto out;
493 }
494
495 csum = nfsd_cache_csum(&rqstp->rq_arg, start, len);
496
497 /*
498 * Since the common case is a cache miss followed by an insert,
499 * preallocate an entry.
500 */
501 rp = nfsd_cacherep_alloc(rqstp, csum, nn);
502 if (!rp)
503 goto out;
504
505 b = nfsd_cache_bucket_find(rqstp->rq_xid, nn);
506 spin_lock(&b->cache_lock);
507 found = nfsd_cache_insert(b, rp, nn);
508 if (found != rp)
509 goto found_entry;
510 *cacherep = rp;
511 rp->c_state = RC_INPROG;
512 nfsd_prune_bucket_locked(nn, b, 3, &dispose);
513 spin_unlock(&b->cache_lock);
514
515 nfsd_cacherep_dispose(&dispose);
516
517 nfsd_stats_rc_misses_inc(nn);
518 atomic_inc(&nn->num_drc_entries);
519 nfsd_stats_drc_mem_usage_add(nn, sizeof(*rp));
520 goto out;
521
522found_entry:
523 /* We found a matching entry which is either in progress or done. */
524 nfsd_reply_cache_free_locked(NULL, rp, nn);
525 nfsd_stats_rc_hits_inc(nn);
526 rtn = RC_DROPIT;
527 rp = found;
528
529 /* Request being processed */
530 if (rp->c_state == RC_INPROG)
531 goto out_trace;
532
533 /* From the hall of fame of impractical attacks:
534 * Is this a user who tries to snoop on the cache? */
535 rtn = RC_DOIT;
536 if (!test_bit(RQ_SECURE, &rqstp->rq_flags) && rp->c_secure)
537 goto out_trace;
538
539 /* Compose RPC reply header */
540 switch (rp->c_type) {
541 case RC_NOCACHE:
542 break;
543 case RC_REPLSTAT:
544 xdr_stream_encode_be32(&rqstp->rq_res_stream, rp->c_replstat);
545 rtn = RC_REPLY;
546 break;
547 case RC_REPLBUFF:
548 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
549 goto out_unlock; /* should not happen */
550 rtn = RC_REPLY;
551 break;
552 default:
553 WARN_ONCE(1, "nfsd: bad repcache type %d\n", rp->c_type);
554 }
555
556out_trace:
557 trace_nfsd_drc_found(nn, rqstp, rtn);
558out_unlock:
559 spin_unlock(&b->cache_lock);
560out:
561 return rtn;
562}
563
564/**
565 * nfsd_cache_update - Update an entry in the duplicate reply cache.
566 * @rqstp: svc_rqst with a finished Reply
567 * @rp: IN: DRC entry for this request
568 * @cachetype: which cache to update
569 * @statp: pointer to Reply's NFS status code, or NULL
570 *
571 * This is called from nfsd_dispatch when the procedure has been
572 * executed and the complete reply is in rqstp->rq_res.
573 *
574 * We're copying around data here rather than swapping buffers because
575 * the toplevel loop requires max-sized buffers, which would be a waste
576 * of memory for a cache with a max reply size of 100 bytes (diropokres).
577 *
578 * If we should start to use different types of cache entries tailored
579 * specifically for attrstat and fh's, we may save even more space.
580 *
581 * Also note that a cachetype of RC_NOCACHE can legally be passed when
582 * nfsd failed to encode a reply that otherwise would have been cached.
583 * In this case, nfsd_cache_update is called with statp == NULL.
584 */
585void nfsd_cache_update(struct svc_rqst *rqstp, struct nfsd_cacherep *rp,
586 int cachetype, __be32 *statp)
587{
588 struct nfsd_net *nn = net_generic(SVC_NET(rqstp), nfsd_net_id);
589 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
590 struct nfsd_drc_bucket *b;
591 int len;
592 size_t bufsize = 0;
593
594 if (!rp)
595 return;
596
597 b = nfsd_cache_bucket_find(rp->c_key.k_xid, nn);
598
599 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
600 len >>= 2;
601
602 /* Don't cache excessive amounts of data and XDR failures */
603 if (!statp || len > (256 >> 2)) {
604 nfsd_reply_cache_free(b, rp, nn);
605 return;
606 }
607
608 switch (cachetype) {
609 case RC_REPLSTAT:
610 if (len != 1)
611 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
612 rp->c_replstat = *statp;
613 break;
614 case RC_REPLBUFF:
615 cachv = &rp->c_replvec;
616 bufsize = len << 2;
617 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
618 if (!cachv->iov_base) {
619 nfsd_reply_cache_free(b, rp, nn);
620 return;
621 }
622 cachv->iov_len = bufsize;
623 memcpy(cachv->iov_base, statp, bufsize);
624 break;
625 case RC_NOCACHE:
626 nfsd_reply_cache_free(b, rp, nn);
627 return;
628 }
629 spin_lock(&b->cache_lock);
630 nfsd_stats_drc_mem_usage_add(nn, bufsize);
631 lru_put_end(b, rp);
632 rp->c_secure = test_bit(RQ_SECURE, &rqstp->rq_flags);
633 rp->c_type = cachetype;
634 rp->c_state = RC_DONE;
635 spin_unlock(&b->cache_lock);
636 return;
637}
638
639static int
640nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
641{
642 __be32 *p;
643
644 p = xdr_reserve_space(&rqstp->rq_res_stream, data->iov_len);
645 if (unlikely(!p))
646 return false;
647 memcpy(p, data->iov_base, data->iov_len);
648 xdr_commit_encode(&rqstp->rq_res_stream);
649 return true;
650}
651
652/*
653 * Note that fields may be added, removed or reordered in the future. Programs
654 * scraping this file for info should test the labels to ensure they're
655 * getting the correct field.
656 */
657int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
658{
659 struct nfsd_net *nn = net_generic(file_inode(m->file)->i_sb->s_fs_info,
660 nfsd_net_id);
661
662 seq_printf(m, "max entries: %u\n", nn->max_drc_entries);
663 seq_printf(m, "num entries: %u\n",
664 atomic_read(&nn->num_drc_entries));
665 seq_printf(m, "hash buckets: %u\n", 1 << nn->maskbits);
666 seq_printf(m, "mem usage: %lld\n",
667 percpu_counter_sum_positive(&nn->counter[NFSD_STATS_DRC_MEM_USAGE]));
668 seq_printf(m, "cache hits: %lld\n",
669 percpu_counter_sum_positive(&nn->counter[NFSD_STATS_RC_HITS]));
670 seq_printf(m, "cache misses: %lld\n",
671 percpu_counter_sum_positive(&nn->counter[NFSD_STATS_RC_MISSES]));
672 seq_printf(m, "not cached: %lld\n",
673 percpu_counter_sum_positive(&nn->counter[NFSD_STATS_RC_NOCACHE]));
674 seq_printf(m, "payload misses: %lld\n",
675 percpu_counter_sum_positive(&nn->counter[NFSD_STATS_PAYLOAD_MISSES]));
676 seq_printf(m, "longest chain len: %u\n", nn->longest_chain);
677 seq_printf(m, "cachesize at longest: %u\n", nn->longest_chain_cachesize);
678 return 0;
679}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Request reply cache. This is currently a global cache, but this may
4 * change in the future and be a per-client cache.
5 *
6 * This code is heavily inspired by the 44BSD implementation, although
7 * it does things a bit differently.
8 *
9 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
10 */
11
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>
19
20#include "nfsd.h"
21#include "cache.h"
22
23#define NFSDDBG_FACILITY NFSDDBG_REPCACHE
24
25/*
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.
29 */
30#define TARGET_BUCKET_SIZE 64
31
32struct nfsd_drc_bucket {
33 struct list_head lru_head;
34 spinlock_t cache_lock;
35};
36
37static struct nfsd_drc_bucket *drc_hashtbl;
38static struct kmem_cache *drc_slab;
39
40/* max number of entries allowed in the cache */
41static unsigned int max_drc_entries;
42
43/* number of significant bits in the hash value */
44static unsigned int maskbits;
45static unsigned int drc_hashsize;
46
47/*
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
50 */
51
52/* total number of entries */
53static atomic_t num_drc_entries;
54
55/* cache misses due only to checksum comparison failures */
56static unsigned int payload_misses;
57
58/* amount of memory (in bytes) currently consumed by the DRC */
59static unsigned int drc_mem_usage;
60
61/* longest hash chain seen */
62static unsigned int longest_chain;
63
64/* size of cache when we saw the longest hash chain */
65static unsigned int longest_chain_cachesize;
66
67static int nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
68static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
69 struct shrink_control *sc);
70static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
71 struct shrink_control *sc);
72
73static struct shrinker nfsd_reply_cache_shrinker = {
74 .scan_objects = nfsd_reply_cache_scan,
75 .count_objects = nfsd_reply_cache_count,
76 .seeks = 1,
77};
78
79/*
80 * Put a cap on the size of the DRC based on the amount of available
81 * low memory in the machine.
82 *
83 * 64MB: 8192
84 * 128MB: 11585
85 * 256MB: 16384
86 * 512MB: 23170
87 * 1GB: 32768
88 * 2GB: 46340
89 * 4GB: 65536
90 * 8GB: 92681
91 * 16GB: 131072
92 *
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
95 * used in k.
96 */
97static unsigned int
98nfsd_cache_size_limit(void)
99{
100 unsigned int limit;
101 unsigned long low_pages = totalram_pages - totalhigh_pages;
102
103 limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
104 return min_t(unsigned int, limit, 256*1024);
105}
106
107/*
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.
110 */
111static unsigned int
112nfsd_hashsize(unsigned int limit)
113{
114 return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
115}
116
117static u32
118nfsd_cache_hash(__be32 xid)
119{
120 return hash_32(be32_to_cpu(xid), maskbits);
121}
122
123static struct svc_cacherep *
124nfsd_reply_cache_alloc(void)
125{
126 struct svc_cacherep *rp;
127
128 rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
129 if (rp) {
130 rp->c_state = RC_UNUSED;
131 rp->c_type = RC_NOCACHE;
132 INIT_LIST_HEAD(&rp->c_lru);
133 }
134 return rp;
135}
136
137static void
138nfsd_reply_cache_free_locked(struct svc_cacherep *rp)
139{
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);
143 }
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);
148}
149
150static void
151nfsd_reply_cache_free(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
152{
153 spin_lock(&b->cache_lock);
154 nfsd_reply_cache_free_locked(rp);
155 spin_unlock(&b->cache_lock);
156}
157
158int nfsd_reply_cache_init(void)
159{
160 unsigned int hashsize;
161 unsigned int i;
162 int status = 0;
163
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);
168
169 status = register_shrinker(&nfsd_reply_cache_shrinker);
170 if (status)
171 return status;
172
173 drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
174 0, 0, NULL);
175 if (!drc_slab)
176 goto out_nomem;
177
178 drc_hashtbl = kcalloc(hashsize, sizeof(*drc_hashtbl), GFP_KERNEL);
179 if (!drc_hashtbl) {
180 drc_hashtbl = vzalloc(hashsize * sizeof(*drc_hashtbl));
181 if (!drc_hashtbl)
182 goto out_nomem;
183 }
184
185 for (i = 0; i < hashsize; i++) {
186 INIT_LIST_HEAD(&drc_hashtbl[i].lru_head);
187 spin_lock_init(&drc_hashtbl[i].cache_lock);
188 }
189 drc_hashsize = hashsize;
190
191 return 0;
192out_nomem:
193 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
194 nfsd_reply_cache_shutdown();
195 return -ENOMEM;
196}
197
198void nfsd_reply_cache_shutdown(void)
199{
200 struct svc_cacherep *rp;
201 unsigned int i;
202
203 unregister_shrinker(&nfsd_reply_cache_shrinker);
204
205 for (i = 0; i < drc_hashsize; i++) {
206 struct list_head *head = &drc_hashtbl[i].lru_head;
207 while (!list_empty(head)) {
208 rp = list_first_entry(head, struct svc_cacherep, c_lru);
209 nfsd_reply_cache_free_locked(rp);
210 }
211 }
212
213 kvfree(drc_hashtbl);
214 drc_hashtbl = NULL;
215 drc_hashsize = 0;
216
217 kmem_cache_destroy(drc_slab);
218 drc_slab = NULL;
219}
220
221/*
222 * Move cache entry to end of LRU list, and queue the cleaner to run if it's
223 * not already scheduled.
224 */
225static void
226lru_put_end(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
227{
228 rp->c_timestamp = jiffies;
229 list_move_tail(&rp->c_lru, &b->lru_head);
230}
231
232static long
233prune_bucket(struct nfsd_drc_bucket *b)
234{
235 struct svc_cacherep *rp, *tmp;
236 long freed = 0;
237
238 list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
239 /*
240 * Don't free entries attached to calls that are still
241 * in-progress, but do keep scanning the list.
242 */
243 if (rp->c_state == RC_INPROG)
244 continue;
245 if (atomic_read(&num_drc_entries) <= max_drc_entries &&
246 time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
247 break;
248 nfsd_reply_cache_free_locked(rp);
249 freed++;
250 }
251 return freed;
252}
253
254/*
255 * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
256 * Also prune the oldest ones when the total exceeds the max number of entries.
257 */
258static long
259prune_cache_entries(void)
260{
261 unsigned int i;
262 long freed = 0;
263
264 for (i = 0; i < drc_hashsize; i++) {
265 struct nfsd_drc_bucket *b = &drc_hashtbl[i];
266
267 if (list_empty(&b->lru_head))
268 continue;
269 spin_lock(&b->cache_lock);
270 freed += prune_bucket(b);
271 spin_unlock(&b->cache_lock);
272 }
273 return freed;
274}
275
276static unsigned long
277nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
278{
279 return atomic_read(&num_drc_entries);
280}
281
282static unsigned long
283nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
284{
285 return prune_cache_entries();
286}
287/*
288 * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
289 */
290static __wsum
291nfsd_cache_csum(struct svc_rqst *rqstp)
292{
293 int idx;
294 unsigned int base;
295 __wsum csum;
296 struct xdr_buf *buf = &rqstp->rq_arg;
297 const unsigned char *p = buf->head[0].iov_base;
298 size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
299 RC_CSUMLEN);
300 size_t len = min(buf->head[0].iov_len, csum_len);
301
302 /* rq_arg.head first */
303 csum = csum_partial(p, len, 0);
304 csum_len -= len;
305
306 /* Continue into page array */
307 idx = buf->page_base / PAGE_SIZE;
308 base = buf->page_base & ~PAGE_MASK;
309 while (csum_len) {
310 p = page_address(buf->pages[idx]) + base;
311 len = min_t(size_t, PAGE_SIZE - base, csum_len);
312 csum = csum_partial(p, len, csum);
313 csum_len -= len;
314 base = 0;
315 ++idx;
316 }
317 return csum;
318}
319
320static bool
321nfsd_cache_match(struct svc_rqst *rqstp, __wsum csum, struct svc_cacherep *rp)
322{
323 /* Check RPC XID first */
324 if (rqstp->rq_xid != rp->c_xid)
325 return false;
326 /* compare checksum of NFS data */
327 if (csum != rp->c_csum) {
328 ++payload_misses;
329 return false;
330 }
331
332 /* Other discriminators */
333 if (rqstp->rq_proc != rp->c_proc ||
334 rqstp->rq_prot != rp->c_prot ||
335 rqstp->rq_vers != rp->c_vers ||
336 rqstp->rq_arg.len != rp->c_len ||
337 !rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) ||
338 rpc_get_port(svc_addr(rqstp)) != rpc_get_port((struct sockaddr *)&rp->c_addr))
339 return false;
340
341 return true;
342}
343
344/*
345 * Search the request hash for an entry that matches the given rqstp.
346 * Must be called with cache_lock held. Returns the found entry or
347 * NULL on failure.
348 */
349static struct svc_cacherep *
350nfsd_cache_search(struct nfsd_drc_bucket *b, struct svc_rqst *rqstp,
351 __wsum csum)
352{
353 struct svc_cacherep *rp, *ret = NULL;
354 struct list_head *rh = &b->lru_head;
355 unsigned int entries = 0;
356
357 list_for_each_entry(rp, rh, c_lru) {
358 ++entries;
359 if (nfsd_cache_match(rqstp, csum, rp)) {
360 ret = rp;
361 break;
362 }
363 }
364
365 /* tally hash chain length stats */
366 if (entries > longest_chain) {
367 longest_chain = entries;
368 longest_chain_cachesize = atomic_read(&num_drc_entries);
369 } else if (entries == longest_chain) {
370 /* prefer to keep the smallest cachesize possible here */
371 longest_chain_cachesize = min_t(unsigned int,
372 longest_chain_cachesize,
373 atomic_read(&num_drc_entries));
374 }
375
376 return ret;
377}
378
379/*
380 * Try to find an entry matching the current call in the cache. When none
381 * is found, we try to grab the oldest expired entry off the LRU list. If
382 * a suitable one isn't there, then drop the cache_lock and allocate a
383 * new one, then search again in case one got inserted while this thread
384 * didn't hold the lock.
385 */
386int
387nfsd_cache_lookup(struct svc_rqst *rqstp)
388{
389 struct svc_cacherep *rp, *found;
390 __be32 xid = rqstp->rq_xid;
391 u32 proto = rqstp->rq_prot,
392 vers = rqstp->rq_vers,
393 proc = rqstp->rq_proc;
394 __wsum csum;
395 u32 hash = nfsd_cache_hash(xid);
396 struct nfsd_drc_bucket *b = &drc_hashtbl[hash];
397 unsigned long age;
398 int type = rqstp->rq_cachetype;
399 int rtn = RC_DOIT;
400
401 rqstp->rq_cacherep = NULL;
402 if (type == RC_NOCACHE) {
403 nfsdstats.rcnocache++;
404 return rtn;
405 }
406
407 csum = nfsd_cache_csum(rqstp);
408
409 /*
410 * Since the common case is a cache miss followed by an insert,
411 * preallocate an entry.
412 */
413 rp = nfsd_reply_cache_alloc();
414 spin_lock(&b->cache_lock);
415 if (likely(rp)) {
416 atomic_inc(&num_drc_entries);
417 drc_mem_usage += sizeof(*rp);
418 }
419
420 /* go ahead and prune the cache */
421 prune_bucket(b);
422
423 found = nfsd_cache_search(b, rqstp, csum);
424 if (found) {
425 if (likely(rp))
426 nfsd_reply_cache_free_locked(rp);
427 rp = found;
428 goto found_entry;
429 }
430
431 if (!rp) {
432 dprintk("nfsd: unable to allocate DRC entry!\n");
433 goto out;
434 }
435
436 nfsdstats.rcmisses++;
437 rqstp->rq_cacherep = rp;
438 rp->c_state = RC_INPROG;
439 rp->c_xid = xid;
440 rp->c_proc = proc;
441 rpc_copy_addr((struct sockaddr *)&rp->c_addr, svc_addr(rqstp));
442 rpc_set_port((struct sockaddr *)&rp->c_addr, rpc_get_port(svc_addr(rqstp)));
443 rp->c_prot = proto;
444 rp->c_vers = vers;
445 rp->c_len = rqstp->rq_arg.len;
446 rp->c_csum = csum;
447
448 lru_put_end(b, rp);
449
450 /* release any buffer */
451 if (rp->c_type == RC_REPLBUFF) {
452 drc_mem_usage -= rp->c_replvec.iov_len;
453 kfree(rp->c_replvec.iov_base);
454 rp->c_replvec.iov_base = NULL;
455 }
456 rp->c_type = RC_NOCACHE;
457 out:
458 spin_unlock(&b->cache_lock);
459 return rtn;
460
461found_entry:
462 nfsdstats.rchits++;
463 /* We found a matching entry which is either in progress or done. */
464 age = jiffies - rp->c_timestamp;
465 lru_put_end(b, rp);
466
467 rtn = RC_DROPIT;
468 /* Request being processed or excessive rexmits */
469 if (rp->c_state == RC_INPROG || age < RC_DELAY)
470 goto out;
471
472 /* From the hall of fame of impractical attacks:
473 * Is this a user who tries to snoop on the cache? */
474 rtn = RC_DOIT;
475 if (!test_bit(RQ_SECURE, &rqstp->rq_flags) && rp->c_secure)
476 goto out;
477
478 /* Compose RPC reply header */
479 switch (rp->c_type) {
480 case RC_NOCACHE:
481 break;
482 case RC_REPLSTAT:
483 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
484 rtn = RC_REPLY;
485 break;
486 case RC_REPLBUFF:
487 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
488 goto out; /* should not happen */
489 rtn = RC_REPLY;
490 break;
491 default:
492 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
493 nfsd_reply_cache_free_locked(rp);
494 }
495
496 goto out;
497}
498
499/*
500 * Update a cache entry. This is called from nfsd_dispatch when
501 * the procedure has been executed and the complete reply is in
502 * rqstp->rq_res.
503 *
504 * We're copying around data here rather than swapping buffers because
505 * the toplevel loop requires max-sized buffers, which would be a waste
506 * of memory for a cache with a max reply size of 100 bytes (diropokres).
507 *
508 * If we should start to use different types of cache entries tailored
509 * specifically for attrstat and fh's, we may save even more space.
510 *
511 * Also note that a cachetype of RC_NOCACHE can legally be passed when
512 * nfsd failed to encode a reply that otherwise would have been cached.
513 * In this case, nfsd_cache_update is called with statp == NULL.
514 */
515void
516nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
517{
518 struct svc_cacherep *rp = rqstp->rq_cacherep;
519 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
520 u32 hash;
521 struct nfsd_drc_bucket *b;
522 int len;
523 size_t bufsize = 0;
524
525 if (!rp)
526 return;
527
528 hash = nfsd_cache_hash(rp->c_xid);
529 b = &drc_hashtbl[hash];
530
531 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
532 len >>= 2;
533
534 /* Don't cache excessive amounts of data and XDR failures */
535 if (!statp || len > (256 >> 2)) {
536 nfsd_reply_cache_free(b, rp);
537 return;
538 }
539
540 switch (cachetype) {
541 case RC_REPLSTAT:
542 if (len != 1)
543 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
544 rp->c_replstat = *statp;
545 break;
546 case RC_REPLBUFF:
547 cachv = &rp->c_replvec;
548 bufsize = len << 2;
549 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
550 if (!cachv->iov_base) {
551 nfsd_reply_cache_free(b, rp);
552 return;
553 }
554 cachv->iov_len = bufsize;
555 memcpy(cachv->iov_base, statp, bufsize);
556 break;
557 case RC_NOCACHE:
558 nfsd_reply_cache_free(b, rp);
559 return;
560 }
561 spin_lock(&b->cache_lock);
562 drc_mem_usage += bufsize;
563 lru_put_end(b, rp);
564 rp->c_secure = test_bit(RQ_SECURE, &rqstp->rq_flags);
565 rp->c_type = cachetype;
566 rp->c_state = RC_DONE;
567 spin_unlock(&b->cache_lock);
568 return;
569}
570
571/*
572 * Copy cached reply to current reply buffer. Should always fit.
573 * FIXME as reply is in a page, we should just attach the page, and
574 * keep a refcount....
575 */
576static int
577nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
578{
579 struct kvec *vec = &rqstp->rq_res.head[0];
580
581 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
582 printk(KERN_WARNING "nfsd: cached reply too large (%zd).\n",
583 data->iov_len);
584 return 0;
585 }
586 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
587 vec->iov_len += data->iov_len;
588 return 1;
589}
590
591/*
592 * Note that fields may be added, removed or reordered in the future. Programs
593 * scraping this file for info should test the labels to ensure they're
594 * getting the correct field.
595 */
596static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
597{
598 seq_printf(m, "max entries: %u\n", max_drc_entries);
599 seq_printf(m, "num entries: %u\n",
600 atomic_read(&num_drc_entries));
601 seq_printf(m, "hash buckets: %u\n", 1 << maskbits);
602 seq_printf(m, "mem usage: %u\n", drc_mem_usage);
603 seq_printf(m, "cache hits: %u\n", nfsdstats.rchits);
604 seq_printf(m, "cache misses: %u\n", nfsdstats.rcmisses);
605 seq_printf(m, "not cached: %u\n", nfsdstats.rcnocache);
606 seq_printf(m, "payload misses: %u\n", payload_misses);
607 seq_printf(m, "longest chain len: %u\n", longest_chain);
608 seq_printf(m, "cachesize at longest: %u\n", longest_chain_cachesize);
609 return 0;
610}
611
612int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
613{
614 return single_open(file, nfsd_reply_cache_stats_show, NULL);
615}