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_create("nfsd_drc",
170 sizeof(struct nfsd_cacherep), 0, 0, NULL);
171 return drc_slab ? 0: -ENOMEM;
172}
173
174void nfsd_drc_slab_free(void)
175{
176 kmem_cache_destroy(drc_slab);
177}
178
179/**
180 * nfsd_net_reply_cache_init - per net namespace reply cache set-up
181 * @nn: nfsd_net being initialized
182 *
183 * Returns zero on succes; otherwise a negative errno is returned.
184 */
185int nfsd_net_reply_cache_init(struct nfsd_net *nn)
186{
187 return nfsd_percpu_counters_init(nn->counter, NFSD_NET_COUNTERS_NUM);
188}
189
190/**
191 * nfsd_net_reply_cache_destroy - per net namespace reply cache tear-down
192 * @nn: nfsd_net being freed
193 *
194 */
195void nfsd_net_reply_cache_destroy(struct nfsd_net *nn)
196{
197 nfsd_percpu_counters_destroy(nn->counter, NFSD_NET_COUNTERS_NUM);
198}
199
200int nfsd_reply_cache_init(struct nfsd_net *nn)
201{
202 unsigned int hashsize;
203 unsigned int i;
204
205 nn->max_drc_entries = nfsd_cache_size_limit();
206 atomic_set(&nn->num_drc_entries, 0);
207 hashsize = nfsd_hashsize(nn->max_drc_entries);
208 nn->maskbits = ilog2(hashsize);
209
210 nn->drc_hashtbl = kvzalloc(array_size(hashsize,
211 sizeof(*nn->drc_hashtbl)), GFP_KERNEL);
212 if (!nn->drc_hashtbl)
213 return -ENOMEM;
214
215 nn->nfsd_reply_cache_shrinker = shrinker_alloc(0, "nfsd-reply:%s",
216 nn->nfsd_name);
217 if (!nn->nfsd_reply_cache_shrinker)
218 goto out_shrinker;
219
220 nn->nfsd_reply_cache_shrinker->scan_objects = nfsd_reply_cache_scan;
221 nn->nfsd_reply_cache_shrinker->count_objects = nfsd_reply_cache_count;
222 nn->nfsd_reply_cache_shrinker->seeks = 1;
223 nn->nfsd_reply_cache_shrinker->private_data = nn;
224
225 shrinker_register(nn->nfsd_reply_cache_shrinker);
226
227 for (i = 0; i < hashsize; i++) {
228 INIT_LIST_HEAD(&nn->drc_hashtbl[i].lru_head);
229 spin_lock_init(&nn->drc_hashtbl[i].cache_lock);
230 }
231 nn->drc_hashsize = hashsize;
232
233 return 0;
234out_shrinker:
235 kvfree(nn->drc_hashtbl);
236 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
237 return -ENOMEM;
238}
239
240void nfsd_reply_cache_shutdown(struct nfsd_net *nn)
241{
242 struct nfsd_cacherep *rp;
243 unsigned int i;
244
245 shrinker_free(nn->nfsd_reply_cache_shrinker);
246
247 for (i = 0; i < nn->drc_hashsize; i++) {
248 struct list_head *head = &nn->drc_hashtbl[i].lru_head;
249 while (!list_empty(head)) {
250 rp = list_first_entry(head, struct nfsd_cacherep, c_lru);
251 nfsd_reply_cache_free_locked(&nn->drc_hashtbl[i],
252 rp, nn);
253 }
254 }
255
256 kvfree(nn->drc_hashtbl);
257 nn->drc_hashtbl = NULL;
258 nn->drc_hashsize = 0;
259
260}
261
262/*
263 * Move cache entry to end of LRU list, and queue the cleaner to run if it's
264 * not already scheduled.
265 */
266static void
267lru_put_end(struct nfsd_drc_bucket *b, struct nfsd_cacherep *rp)
268{
269 rp->c_timestamp = jiffies;
270 list_move_tail(&rp->c_lru, &b->lru_head);
271}
272
273static noinline struct nfsd_drc_bucket *
274nfsd_cache_bucket_find(__be32 xid, struct nfsd_net *nn)
275{
276 unsigned int hash = hash_32((__force u32)xid, nn->maskbits);
277
278 return &nn->drc_hashtbl[hash];
279}
280
281/*
282 * Remove and return no more than @max expired entries in bucket @b.
283 * If @max is zero, do not limit the number of removed entries.
284 */
285static void
286nfsd_prune_bucket_locked(struct nfsd_net *nn, struct nfsd_drc_bucket *b,
287 unsigned int max, struct list_head *dispose)
288{
289 unsigned long expiry = jiffies - RC_EXPIRE;
290 struct nfsd_cacherep *rp, *tmp;
291 unsigned int freed = 0;
292
293 lockdep_assert_held(&b->cache_lock);
294
295 /* The bucket LRU is ordered oldest-first. */
296 list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
297 /*
298 * Don't free entries attached to calls that are still
299 * in-progress, but do keep scanning the list.
300 */
301 if (rp->c_state == RC_INPROG)
302 continue;
303
304 if (atomic_read(&nn->num_drc_entries) <= nn->max_drc_entries &&
305 time_before(expiry, rp->c_timestamp))
306 break;
307
308 nfsd_cacherep_unlink_locked(nn, b, rp);
309 list_add(&rp->c_lru, dispose);
310
311 if (max && ++freed > max)
312 break;
313 }
314}
315
316/**
317 * nfsd_reply_cache_count - count_objects method for the DRC shrinker
318 * @shrink: our registered shrinker context
319 * @sc: garbage collection parameters
320 *
321 * Returns the total number of entries in the duplicate reply cache. To
322 * keep things simple and quick, this is not the number of expired entries
323 * in the cache (ie, the number that would be removed by a call to
324 * nfsd_reply_cache_scan).
325 */
326static unsigned long
327nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
328{
329 struct nfsd_net *nn = shrink->private_data;
330
331 return atomic_read(&nn->num_drc_entries);
332}
333
334/**
335 * nfsd_reply_cache_scan - scan_objects method for the DRC shrinker
336 * @shrink: our registered shrinker context
337 * @sc: garbage collection parameters
338 *
339 * Free expired entries on each bucket's LRU list until we've released
340 * nr_to_scan freed objects. Nothing will be released if the cache
341 * has not exceeded it's max_drc_entries limit.
342 *
343 * Returns the number of entries released by this call.
344 */
345static unsigned long
346nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
347{
348 struct nfsd_net *nn = shrink->private_data;
349 unsigned long freed = 0;
350 LIST_HEAD(dispose);
351 unsigned int i;
352
353 for (i = 0; i < nn->drc_hashsize; i++) {
354 struct nfsd_drc_bucket *b = &nn->drc_hashtbl[i];
355
356 if (list_empty(&b->lru_head))
357 continue;
358
359 spin_lock(&b->cache_lock);
360 nfsd_prune_bucket_locked(nn, b, 0, &dispose);
361 spin_unlock(&b->cache_lock);
362
363 freed += nfsd_cacherep_dispose(&dispose);
364 if (freed > sc->nr_to_scan)
365 break;
366 }
367 return freed;
368}
369
370/**
371 * nfsd_cache_csum - Checksum incoming NFS Call arguments
372 * @buf: buffer containing a whole RPC Call message
373 * @start: starting byte of the NFS Call header
374 * @remaining: size of the NFS Call header, in bytes
375 *
376 * Compute a weak checksum of the leading bytes of an NFS procedure
377 * call header to help verify that a retransmitted Call matches an
378 * entry in the duplicate reply cache.
379 *
380 * To avoid assumptions about how the RPC message is laid out in
381 * @buf and what else it might contain (eg, a GSS MIC suffix), the
382 * caller passes us the exact location and length of the NFS Call
383 * header.
384 *
385 * Returns a 32-bit checksum value, as defined in RFC 793.
386 */
387static __wsum nfsd_cache_csum(struct xdr_buf *buf, unsigned int start,
388 unsigned int remaining)
389{
390 unsigned int base, len;
391 struct xdr_buf subbuf;
392 __wsum csum = 0;
393 void *p;
394 int idx;
395
396 if (remaining > RC_CSUMLEN)
397 remaining = RC_CSUMLEN;
398 if (xdr_buf_subsegment(buf, &subbuf, start, remaining))
399 return csum;
400
401 /* rq_arg.head first */
402 if (subbuf.head[0].iov_len) {
403 len = min_t(unsigned int, subbuf.head[0].iov_len, remaining);
404 csum = csum_partial(subbuf.head[0].iov_base, len, csum);
405 remaining -= len;
406 }
407
408 /* Continue into page array */
409 idx = subbuf.page_base / PAGE_SIZE;
410 base = subbuf.page_base & ~PAGE_MASK;
411 while (remaining) {
412 p = page_address(subbuf.pages[idx]) + base;
413 len = min_t(unsigned int, PAGE_SIZE - base, remaining);
414 csum = csum_partial(p, len, csum);
415 remaining -= len;
416 base = 0;
417 ++idx;
418 }
419 return csum;
420}
421
422static int
423nfsd_cache_key_cmp(const struct nfsd_cacherep *key,
424 const struct nfsd_cacherep *rp, struct nfsd_net *nn)
425{
426 if (key->c_key.k_xid == rp->c_key.k_xid &&
427 key->c_key.k_csum != rp->c_key.k_csum) {
428 nfsd_stats_payload_misses_inc(nn);
429 trace_nfsd_drc_mismatch(nn, key, rp);
430 }
431
432 return memcmp(&key->c_key, &rp->c_key, sizeof(key->c_key));
433}
434
435/*
436 * Search the request hash for an entry that matches the given rqstp.
437 * Must be called with cache_lock held. Returns the found entry or
438 * inserts an empty key on failure.
439 */
440static struct nfsd_cacherep *
441nfsd_cache_insert(struct nfsd_drc_bucket *b, struct nfsd_cacherep *key,
442 struct nfsd_net *nn)
443{
444 struct nfsd_cacherep *rp, *ret = key;
445 struct rb_node **p = &b->rb_head.rb_node,
446 *parent = NULL;
447 unsigned int entries = 0;
448 int cmp;
449
450 while (*p != NULL) {
451 ++entries;
452 parent = *p;
453 rp = rb_entry(parent, struct nfsd_cacherep, c_node);
454
455 cmp = nfsd_cache_key_cmp(key, rp, nn);
456 if (cmp < 0)
457 p = &parent->rb_left;
458 else if (cmp > 0)
459 p = &parent->rb_right;
460 else {
461 ret = rp;
462 goto out;
463 }
464 }
465 rb_link_node(&key->c_node, parent, p);
466 rb_insert_color(&key->c_node, &b->rb_head);
467out:
468 /* tally hash chain length stats */
469 if (entries > nn->longest_chain) {
470 nn->longest_chain = entries;
471 nn->longest_chain_cachesize = atomic_read(&nn->num_drc_entries);
472 } else if (entries == nn->longest_chain) {
473 /* prefer to keep the smallest cachesize possible here */
474 nn->longest_chain_cachesize = min_t(unsigned int,
475 nn->longest_chain_cachesize,
476 atomic_read(&nn->num_drc_entries));
477 }
478
479 lru_put_end(b, ret);
480 return ret;
481}
482
483/**
484 * nfsd_cache_lookup - Find an entry in the duplicate reply cache
485 * @rqstp: Incoming Call to find
486 * @start: starting byte in @rqstp->rq_arg of the NFS Call header
487 * @len: size of the NFS Call header, in bytes
488 * @cacherep: OUT: DRC entry for this request
489 *
490 * Try to find an entry matching the current call in the cache. When none
491 * is found, we try to grab the oldest expired entry off the LRU list. If
492 * a suitable one isn't there, then drop the cache_lock and allocate a
493 * new one, then search again in case one got inserted while this thread
494 * didn't hold the lock.
495 *
496 * Return values:
497 * %RC_DOIT: Process the request normally
498 * %RC_REPLY: Reply from cache
499 * %RC_DROPIT: Do not process the request further
500 */
501int nfsd_cache_lookup(struct svc_rqst *rqstp, unsigned int start,
502 unsigned int len, struct nfsd_cacherep **cacherep)
503{
504 struct nfsd_net *nn;
505 struct nfsd_cacherep *rp, *found;
506 __wsum csum;
507 struct nfsd_drc_bucket *b;
508 int type = rqstp->rq_cachetype;
509 LIST_HEAD(dispose);
510 int rtn = RC_DOIT;
511
512 if (type == RC_NOCACHE) {
513 nfsd_stats_rc_nocache_inc();
514 goto out;
515 }
516
517 csum = nfsd_cache_csum(&rqstp->rq_arg, start, len);
518
519 /*
520 * Since the common case is a cache miss followed by an insert,
521 * preallocate an entry.
522 */
523 nn = net_generic(SVC_NET(rqstp), nfsd_net_id);
524 rp = nfsd_cacherep_alloc(rqstp, csum, nn);
525 if (!rp)
526 goto out;
527
528 b = nfsd_cache_bucket_find(rqstp->rq_xid, nn);
529 spin_lock(&b->cache_lock);
530 found = nfsd_cache_insert(b, rp, nn);
531 if (found != rp)
532 goto found_entry;
533 *cacherep = rp;
534 rp->c_state = RC_INPROG;
535 nfsd_prune_bucket_locked(nn, b, 3, &dispose);
536 spin_unlock(&b->cache_lock);
537
538 nfsd_cacherep_dispose(&dispose);
539
540 nfsd_stats_rc_misses_inc();
541 atomic_inc(&nn->num_drc_entries);
542 nfsd_stats_drc_mem_usage_add(nn, sizeof(*rp));
543 goto out;
544
545found_entry:
546 /* We found a matching entry which is either in progress or done. */
547 nfsd_reply_cache_free_locked(NULL, rp, nn);
548 nfsd_stats_rc_hits_inc();
549 rtn = RC_DROPIT;
550 rp = found;
551
552 /* Request being processed */
553 if (rp->c_state == RC_INPROG)
554 goto out_trace;
555
556 /* From the hall of fame of impractical attacks:
557 * Is this a user who tries to snoop on the cache? */
558 rtn = RC_DOIT;
559 if (!test_bit(RQ_SECURE, &rqstp->rq_flags) && rp->c_secure)
560 goto out_trace;
561
562 /* Compose RPC reply header */
563 switch (rp->c_type) {
564 case RC_NOCACHE:
565 break;
566 case RC_REPLSTAT:
567 xdr_stream_encode_be32(&rqstp->rq_res_stream, rp->c_replstat);
568 rtn = RC_REPLY;
569 break;
570 case RC_REPLBUFF:
571 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
572 goto out_unlock; /* should not happen */
573 rtn = RC_REPLY;
574 break;
575 default:
576 WARN_ONCE(1, "nfsd: bad repcache type %d\n", rp->c_type);
577 }
578
579out_trace:
580 trace_nfsd_drc_found(nn, rqstp, rtn);
581out_unlock:
582 spin_unlock(&b->cache_lock);
583out:
584 return rtn;
585}
586
587/**
588 * nfsd_cache_update - Update an entry in the duplicate reply cache.
589 * @rqstp: svc_rqst with a finished Reply
590 * @rp: IN: DRC entry for this request
591 * @cachetype: which cache to update
592 * @statp: pointer to Reply's NFS status code, or NULL
593 *
594 * This is called from nfsd_dispatch when the procedure has been
595 * executed and the complete reply is in rqstp->rq_res.
596 *
597 * We're copying around data here rather than swapping buffers because
598 * the toplevel loop requires max-sized buffers, which would be a waste
599 * of memory for a cache with a max reply size of 100 bytes (diropokres).
600 *
601 * If we should start to use different types of cache entries tailored
602 * specifically for attrstat and fh's, we may save even more space.
603 *
604 * Also note that a cachetype of RC_NOCACHE can legally be passed when
605 * nfsd failed to encode a reply that otherwise would have been cached.
606 * In this case, nfsd_cache_update is called with statp == NULL.
607 */
608void nfsd_cache_update(struct svc_rqst *rqstp, struct nfsd_cacherep *rp,
609 int cachetype, __be32 *statp)
610{
611 struct nfsd_net *nn = net_generic(SVC_NET(rqstp), nfsd_net_id);
612 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
613 struct nfsd_drc_bucket *b;
614 int len;
615 size_t bufsize = 0;
616
617 if (!rp)
618 return;
619
620 b = nfsd_cache_bucket_find(rp->c_key.k_xid, nn);
621
622 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
623 len >>= 2;
624
625 /* Don't cache excessive amounts of data and XDR failures */
626 if (!statp || len > (256 >> 2)) {
627 nfsd_reply_cache_free(b, rp, nn);
628 return;
629 }
630
631 switch (cachetype) {
632 case RC_REPLSTAT:
633 if (len != 1)
634 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
635 rp->c_replstat = *statp;
636 break;
637 case RC_REPLBUFF:
638 cachv = &rp->c_replvec;
639 bufsize = len << 2;
640 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
641 if (!cachv->iov_base) {
642 nfsd_reply_cache_free(b, rp, nn);
643 return;
644 }
645 cachv->iov_len = bufsize;
646 memcpy(cachv->iov_base, statp, bufsize);
647 break;
648 case RC_NOCACHE:
649 nfsd_reply_cache_free(b, rp, nn);
650 return;
651 }
652 spin_lock(&b->cache_lock);
653 nfsd_stats_drc_mem_usage_add(nn, bufsize);
654 lru_put_end(b, rp);
655 rp->c_secure = test_bit(RQ_SECURE, &rqstp->rq_flags);
656 rp->c_type = cachetype;
657 rp->c_state = RC_DONE;
658 spin_unlock(&b->cache_lock);
659 return;
660}
661
662static int
663nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
664{
665 __be32 *p;
666
667 p = xdr_reserve_space(&rqstp->rq_res_stream, data->iov_len);
668 if (unlikely(!p))
669 return false;
670 memcpy(p, data->iov_base, data->iov_len);
671 xdr_commit_encode(&rqstp->rq_res_stream);
672 return true;
673}
674
675/*
676 * Note that fields may be added, removed or reordered in the future. Programs
677 * scraping this file for info should test the labels to ensure they're
678 * getting the correct field.
679 */
680int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
681{
682 struct nfsd_net *nn = net_generic(file_inode(m->file)->i_sb->s_fs_info,
683 nfsd_net_id);
684
685 seq_printf(m, "max entries: %u\n", nn->max_drc_entries);
686 seq_printf(m, "num entries: %u\n",
687 atomic_read(&nn->num_drc_entries));
688 seq_printf(m, "hash buckets: %u\n", 1 << nn->maskbits);
689 seq_printf(m, "mem usage: %lld\n",
690 percpu_counter_sum_positive(&nn->counter[NFSD_NET_DRC_MEM_USAGE]));
691 seq_printf(m, "cache hits: %lld\n",
692 percpu_counter_sum_positive(&nfsdstats.counter[NFSD_STATS_RC_HITS]));
693 seq_printf(m, "cache misses: %lld\n",
694 percpu_counter_sum_positive(&nfsdstats.counter[NFSD_STATS_RC_MISSES]));
695 seq_printf(m, "not cached: %lld\n",
696 percpu_counter_sum_positive(&nfsdstats.counter[NFSD_STATS_RC_NOCACHE]));
697 seq_printf(m, "payload misses: %lld\n",
698 percpu_counter_sum_positive(&nn->counter[NFSD_NET_PAYLOAD_MISSES]));
699 seq_printf(m, "longest chain len: %u\n", nn->longest_chain);
700 seq_printf(m, "cachesize at longest: %u\n", nn->longest_chain_cachesize);
701 return 0;
702}
1/*
2 * Request reply cache. This is currently a global cache, but this may
3 * change in the future and be a per-client cache.
4 *
5 * This code is heavily inspired by the 44BSD implementation, although
6 * it does things a bit differently.
7 *
8 * Copyright (C) 1995, 1996 Olaf Kirch <okir@monad.swb.de>
9 */
10
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>
17
18#include "nfsd.h"
19#include "cache.h"
20
21#define NFSDDBG_FACILITY NFSDDBG_REPCACHE
22
23/*
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.
27 */
28#define TARGET_BUCKET_SIZE 64
29
30static struct hlist_head * cache_hash;
31static struct list_head lru_head;
32static struct kmem_cache *drc_slab;
33
34/* max number of entries allowed in the cache */
35static unsigned int max_drc_entries;
36
37/* number of significant bits in the hash value */
38static unsigned int maskbits;
39
40/*
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
43 */
44
45/* total number of entries */
46static unsigned int num_drc_entries;
47
48/* cache misses due only to checksum comparison failures */
49static unsigned int payload_misses;
50
51/* amount of memory (in bytes) currently consumed by the DRC */
52static unsigned int drc_mem_usage;
53
54/* longest hash chain seen */
55static unsigned int longest_chain;
56
57/* size of cache when we saw the longest hash chain */
58static unsigned int longest_chain_cachesize;
59
60static int nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
61static void cache_cleaner_func(struct work_struct *unused);
62static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
63 struct shrink_control *sc);
64static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
65 struct shrink_control *sc);
66
67static struct shrinker nfsd_reply_cache_shrinker = {
68 .scan_objects = nfsd_reply_cache_scan,
69 .count_objects = nfsd_reply_cache_count,
70 .seeks = 1,
71};
72
73/*
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.
77 */
78static DEFINE_SPINLOCK(cache_lock);
79static DECLARE_DELAYED_WORK(cache_cleaner, cache_cleaner_func);
80
81/*
82 * Put a cap on the size of the DRC based on the amount of available
83 * low memory in the machine.
84 *
85 * 64MB: 8192
86 * 128MB: 11585
87 * 256MB: 16384
88 * 512MB: 23170
89 * 1GB: 32768
90 * 2GB: 46340
91 * 4GB: 65536
92 * 8GB: 92681
93 * 16GB: 131072
94 *
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
97 * used in k.
98 */
99static unsigned int
100nfsd_cache_size_limit(void)
101{
102 unsigned int limit;
103 unsigned long low_pages = totalram_pages - totalhigh_pages;
104
105 limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
106 return min_t(unsigned int, limit, 256*1024);
107}
108
109/*
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.
112 */
113static unsigned int
114nfsd_hashsize(unsigned int limit)
115{
116 return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
117}
118
119static struct svc_cacherep *
120nfsd_reply_cache_alloc(void)
121{
122 struct svc_cacherep *rp;
123
124 rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
125 if (rp) {
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);
130 }
131 return rp;
132}
133
134static void
135nfsd_reply_cache_free_locked(struct svc_cacherep *rp)
136{
137 if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
138 drc_mem_usage -= rp->c_replvec.iov_len;
139 kfree(rp->c_replvec.iov_base);
140 }
141 if (!hlist_unhashed(&rp->c_hash))
142 hlist_del(&rp->c_hash);
143 list_del(&rp->c_lru);
144 --num_drc_entries;
145 drc_mem_usage -= sizeof(*rp);
146 kmem_cache_free(drc_slab, rp);
147}
148
149static void
150nfsd_reply_cache_free(struct svc_cacherep *rp)
151{
152 spin_lock(&cache_lock);
153 nfsd_reply_cache_free_locked(rp);
154 spin_unlock(&cache_lock);
155}
156
157int nfsd_reply_cache_init(void)
158{
159 unsigned int hashsize;
160
161 INIT_LIST_HEAD(&lru_head);
162 max_drc_entries = nfsd_cache_size_limit();
163 num_drc_entries = 0;
164 hashsize = nfsd_hashsize(max_drc_entries);
165 maskbits = ilog2(hashsize);
166
167 register_shrinker(&nfsd_reply_cache_shrinker);
168 drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
169 0, 0, NULL);
170 if (!drc_slab)
171 goto out_nomem;
172
173 cache_hash = kcalloc(hashsize, sizeof(struct hlist_head), GFP_KERNEL);
174 if (!cache_hash)
175 goto out_nomem;
176
177 return 0;
178out_nomem:
179 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
180 nfsd_reply_cache_shutdown();
181 return -ENOMEM;
182}
183
184void nfsd_reply_cache_shutdown(void)
185{
186 struct svc_cacherep *rp;
187
188 unregister_shrinker(&nfsd_reply_cache_shrinker);
189 cancel_delayed_work_sync(&cache_cleaner);
190
191 while (!list_empty(&lru_head)) {
192 rp = list_entry(lru_head.next, struct svc_cacherep, c_lru);
193 nfsd_reply_cache_free_locked(rp);
194 }
195
196 kfree (cache_hash);
197 cache_hash = NULL;
198
199 if (drc_slab) {
200 kmem_cache_destroy(drc_slab);
201 drc_slab = NULL;
202 }
203}
204
205/*
206 * Move cache entry to end of LRU list, and queue the cleaner to run if it's
207 * not already scheduled.
208 */
209static void
210lru_put_end(struct svc_cacherep *rp)
211{
212 rp->c_timestamp = jiffies;
213 list_move_tail(&rp->c_lru, &lru_head);
214 schedule_delayed_work(&cache_cleaner, RC_EXPIRE);
215}
216
217/*
218 * Move a cache entry from one hash list to another
219 */
220static void
221hash_refile(struct svc_cacherep *rp)
222{
223 hlist_del_init(&rp->c_hash);
224 hlist_add_head(&rp->c_hash, cache_hash + hash_32(rp->c_xid, maskbits));
225}
226
227static inline bool
228nfsd_cache_entry_expired(struct svc_cacherep *rp)
229{
230 return rp->c_state != RC_INPROG &&
231 time_after(jiffies, rp->c_timestamp + RC_EXPIRE);
232}
233
234/*
235 * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
236 * Also prune the oldest ones when the total exceeds the max number of entries.
237 */
238static long
239prune_cache_entries(void)
240{
241 struct svc_cacherep *rp, *tmp;
242 long freed = 0;
243
244 list_for_each_entry_safe(rp, tmp, &lru_head, c_lru) {
245 if (!nfsd_cache_entry_expired(rp) &&
246 num_drc_entries <= max_drc_entries)
247 break;
248 nfsd_reply_cache_free_locked(rp);
249 freed++;
250 }
251
252 /*
253 * Conditionally rearm the job. If we cleaned out the list, then
254 * cancel any pending run (since there won't be any work to do).
255 * Otherwise, we rearm the job or modify the existing one to run in
256 * RC_EXPIRE since we just ran the pruner.
257 */
258 if (list_empty(&lru_head))
259 cancel_delayed_work(&cache_cleaner);
260 else
261 mod_delayed_work(system_wq, &cache_cleaner, RC_EXPIRE);
262 return freed;
263}
264
265static void
266cache_cleaner_func(struct work_struct *unused)
267{
268 spin_lock(&cache_lock);
269 prune_cache_entries();
270 spin_unlock(&cache_lock);
271}
272
273static unsigned long
274nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
275{
276 unsigned long num;
277
278 spin_lock(&cache_lock);
279 num = num_drc_entries;
280 spin_unlock(&cache_lock);
281
282 return num;
283}
284
285static unsigned long
286nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
287{
288 unsigned long freed;
289
290 spin_lock(&cache_lock);
291 freed = prune_cache_entries();
292 spin_unlock(&cache_lock);
293 return freed;
294}
295/*
296 * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
297 */
298static __wsum
299nfsd_cache_csum(struct svc_rqst *rqstp)
300{
301 int idx;
302 unsigned int base;
303 __wsum csum;
304 struct xdr_buf *buf = &rqstp->rq_arg;
305 const unsigned char *p = buf->head[0].iov_base;
306 size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
307 RC_CSUMLEN);
308 size_t len = min(buf->head[0].iov_len, csum_len);
309
310 /* rq_arg.head first */
311 csum = csum_partial(p, len, 0);
312 csum_len -= len;
313
314 /* Continue into page array */
315 idx = buf->page_base / PAGE_SIZE;
316 base = buf->page_base & ~PAGE_MASK;
317 while (csum_len) {
318 p = page_address(buf->pages[idx]) + base;
319 len = min_t(size_t, PAGE_SIZE - base, csum_len);
320 csum = csum_partial(p, len, csum);
321 csum_len -= len;
322 base = 0;
323 ++idx;
324 }
325 return csum;
326}
327
328static bool
329nfsd_cache_match(struct svc_rqst *rqstp, __wsum csum, struct svc_cacherep *rp)
330{
331 /* Check RPC header info first */
332 if (rqstp->rq_xid != rp->c_xid || rqstp->rq_proc != rp->c_proc ||
333 rqstp->rq_prot != rp->c_prot || rqstp->rq_vers != rp->c_vers ||
334 rqstp->rq_arg.len != rp->c_len ||
335 !rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) ||
336 rpc_get_port(svc_addr(rqstp)) != rpc_get_port((struct sockaddr *)&rp->c_addr))
337 return false;
338
339 /* compare checksum of NFS data */
340 if (csum != rp->c_csum) {
341 ++payload_misses;
342 return false;
343 }
344
345 return true;
346}
347
348/*
349 * Search the request hash for an entry that matches the given rqstp.
350 * Must be called with cache_lock held. Returns the found entry or
351 * NULL on failure.
352 */
353static struct svc_cacherep *
354nfsd_cache_search(struct svc_rqst *rqstp, __wsum csum)
355{
356 struct svc_cacherep *rp, *ret = NULL;
357 struct hlist_head *rh;
358 unsigned int entries = 0;
359
360 rh = &cache_hash[hash_32(rqstp->rq_xid, maskbits)];
361 hlist_for_each_entry(rp, rh, c_hash) {
362 ++entries;
363 if (nfsd_cache_match(rqstp, csum, rp)) {
364 ret = rp;
365 break;
366 }
367 }
368
369 /* tally hash chain length stats */
370 if (entries > longest_chain) {
371 longest_chain = entries;
372 longest_chain_cachesize = num_drc_entries;
373 } else if (entries == longest_chain) {
374 /* prefer to keep the smallest cachesize possible here */
375 longest_chain_cachesize = min(longest_chain_cachesize,
376 num_drc_entries);
377 }
378
379 return ret;
380}
381
382/*
383 * Try to find an entry matching the current call in the cache. When none
384 * is found, we try to grab the oldest expired entry off the LRU list. If
385 * a suitable one isn't there, then drop the cache_lock and allocate a
386 * new one, then search again in case one got inserted while this thread
387 * didn't hold the lock.
388 */
389int
390nfsd_cache_lookup(struct svc_rqst *rqstp)
391{
392 struct svc_cacherep *rp, *found;
393 __be32 xid = rqstp->rq_xid;
394 u32 proto = rqstp->rq_prot,
395 vers = rqstp->rq_vers,
396 proc = rqstp->rq_proc;
397 __wsum csum;
398 unsigned long age;
399 int type = rqstp->rq_cachetype;
400 int rtn = RC_DOIT;
401
402 rqstp->rq_cacherep = NULL;
403 if (type == RC_NOCACHE) {
404 nfsdstats.rcnocache++;
405 return rtn;
406 }
407
408 csum = nfsd_cache_csum(rqstp);
409
410 /*
411 * Since the common case is a cache miss followed by an insert,
412 * preallocate an entry.
413 */
414 rp = nfsd_reply_cache_alloc();
415 spin_lock(&cache_lock);
416 if (likely(rp)) {
417 ++num_drc_entries;
418 drc_mem_usage += sizeof(*rp);
419 }
420
421 /* go ahead and prune the cache */
422 prune_cache_entries();
423
424 found = nfsd_cache_search(rqstp, csum);
425 if (found) {
426 if (likely(rp))
427 nfsd_reply_cache_free_locked(rp);
428 rp = found;
429 goto found_entry;
430 }
431
432 if (!rp) {
433 dprintk("nfsd: unable to allocate DRC entry!\n");
434 goto out;
435 }
436
437 nfsdstats.rcmisses++;
438 rqstp->rq_cacherep = rp;
439 rp->c_state = RC_INPROG;
440 rp->c_xid = xid;
441 rp->c_proc = proc;
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)));
444 rp->c_prot = proto;
445 rp->c_vers = vers;
446 rp->c_len = rqstp->rq_arg.len;
447 rp->c_csum = csum;
448
449 hash_refile(rp);
450 lru_put_end(rp);
451
452 /* release any buffer */
453 if (rp->c_type == RC_REPLBUFF) {
454 drc_mem_usage -= rp->c_replvec.iov_len;
455 kfree(rp->c_replvec.iov_base);
456 rp->c_replvec.iov_base = NULL;
457 }
458 rp->c_type = RC_NOCACHE;
459 out:
460 spin_unlock(&cache_lock);
461 return rtn;
462
463found_entry:
464 nfsdstats.rchits++;
465 /* We found a matching entry which is either in progress or done. */
466 age = jiffies - rp->c_timestamp;
467 lru_put_end(rp);
468
469 rtn = RC_DROPIT;
470 /* Request being processed or excessive rexmits */
471 if (rp->c_state == RC_INPROG || age < RC_DELAY)
472 goto out;
473
474 /* From the hall of fame of impractical attacks:
475 * Is this a user who tries to snoop on the cache? */
476 rtn = RC_DOIT;
477 if (!rqstp->rq_secure && rp->c_secure)
478 goto out;
479
480 /* Compose RPC reply header */
481 switch (rp->c_type) {
482 case RC_NOCACHE:
483 break;
484 case RC_REPLSTAT:
485 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
486 rtn = RC_REPLY;
487 break;
488 case RC_REPLBUFF:
489 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
490 goto out; /* should not happen */
491 rtn = RC_REPLY;
492 break;
493 default:
494 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
495 nfsd_reply_cache_free_locked(rp);
496 }
497
498 goto out;
499}
500
501/*
502 * Update a cache entry. This is called from nfsd_dispatch when
503 * the procedure has been executed and the complete reply is in
504 * rqstp->rq_res.
505 *
506 * We're copying around data here rather than swapping buffers because
507 * the toplevel loop requires max-sized buffers, which would be a waste
508 * of memory for a cache with a max reply size of 100 bytes (diropokres).
509 *
510 * If we should start to use different types of cache entries tailored
511 * specifically for attrstat and fh's, we may save even more space.
512 *
513 * Also note that a cachetype of RC_NOCACHE can legally be passed when
514 * nfsd failed to encode a reply that otherwise would have been cached.
515 * In this case, nfsd_cache_update is called with statp == NULL.
516 */
517void
518nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
519{
520 struct svc_cacherep *rp = rqstp->rq_cacherep;
521 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
522 int len;
523 size_t bufsize = 0;
524
525 if (!rp)
526 return;
527
528 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
529 len >>= 2;
530
531 /* Don't cache excessive amounts of data and XDR failures */
532 if (!statp || len > (256 >> 2)) {
533 nfsd_reply_cache_free(rp);
534 return;
535 }
536
537 switch (cachetype) {
538 case RC_REPLSTAT:
539 if (len != 1)
540 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
541 rp->c_replstat = *statp;
542 break;
543 case RC_REPLBUFF:
544 cachv = &rp->c_replvec;
545 bufsize = len << 2;
546 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
547 if (!cachv->iov_base) {
548 nfsd_reply_cache_free(rp);
549 return;
550 }
551 cachv->iov_len = bufsize;
552 memcpy(cachv->iov_base, statp, bufsize);
553 break;
554 case RC_NOCACHE:
555 nfsd_reply_cache_free(rp);
556 return;
557 }
558 spin_lock(&cache_lock);
559 drc_mem_usage += bufsize;
560 lru_put_end(rp);
561 rp->c_secure = rqstp->rq_secure;
562 rp->c_type = cachetype;
563 rp->c_state = RC_DONE;
564 spin_unlock(&cache_lock);
565 return;
566}
567
568/*
569 * Copy cached reply to current reply buffer. Should always fit.
570 * FIXME as reply is in a page, we should just attach the page, and
571 * keep a refcount....
572 */
573static int
574nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
575{
576 struct kvec *vec = &rqstp->rq_res.head[0];
577
578 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
579 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
580 data->iov_len);
581 return 0;
582 }
583 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
584 vec->iov_len += data->iov_len;
585 return 1;
586}
587
588/*
589 * Note that fields may be added, removed or reordered in the future. Programs
590 * scraping this file for info should test the labels to ensure they're
591 * getting the correct field.
592 */
593static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
594{
595 spin_lock(&cache_lock);
596 seq_printf(m, "max entries: %u\n", max_drc_entries);
597 seq_printf(m, "num entries: %u\n", num_drc_entries);
598 seq_printf(m, "hash buckets: %u\n", 1 << maskbits);
599 seq_printf(m, "mem usage: %u\n", drc_mem_usage);
600 seq_printf(m, "cache hits: %u\n", nfsdstats.rchits);
601 seq_printf(m, "cache misses: %u\n", nfsdstats.rcmisses);
602 seq_printf(m, "not cached: %u\n", nfsdstats.rcnocache);
603 seq_printf(m, "payload misses: %u\n", payload_misses);
604 seq_printf(m, "longest chain len: %u\n", longest_chain);
605 seq_printf(m, "cachesize at longest: %u\n", longest_chain_cachesize);
606 spin_unlock(&cache_lock);
607 return 0;
608}
609
610int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
611{
612 return single_open(file, nfsd_reply_cache_stats_show, NULL);
613}