Loading...
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}
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 u32
88nfsd_cache_hash(__be32 xid, struct nfsd_net *nn)
89{
90 return hash_32(be32_to_cpu(xid), nn->maskbits);
91}
92
93static struct svc_cacherep *
94nfsd_reply_cache_alloc(struct svc_rqst *rqstp, __wsum csum,
95 struct nfsd_net *nn)
96{
97 struct svc_cacherep *rp;
98
99 rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
100 if (rp) {
101 rp->c_state = RC_UNUSED;
102 rp->c_type = RC_NOCACHE;
103 RB_CLEAR_NODE(&rp->c_node);
104 INIT_LIST_HEAD(&rp->c_lru);
105
106 memset(&rp->c_key, 0, sizeof(rp->c_key));
107 rp->c_key.k_xid = rqstp->rq_xid;
108 rp->c_key.k_proc = rqstp->rq_proc;
109 rpc_copy_addr((struct sockaddr *)&rp->c_key.k_addr, svc_addr(rqstp));
110 rpc_set_port((struct sockaddr *)&rp->c_key.k_addr, rpc_get_port(svc_addr(rqstp)));
111 rp->c_key.k_prot = rqstp->rq_prot;
112 rp->c_key.k_vers = rqstp->rq_vers;
113 rp->c_key.k_len = rqstp->rq_arg.len;
114 rp->c_key.k_csum = csum;
115 }
116 return rp;
117}
118
119static void
120nfsd_reply_cache_free_locked(struct nfsd_drc_bucket *b, struct svc_cacherep *rp,
121 struct nfsd_net *nn)
122{
123 if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
124 nfsd_stats_drc_mem_usage_sub(nn, rp->c_replvec.iov_len);
125 kfree(rp->c_replvec.iov_base);
126 }
127 if (rp->c_state != RC_UNUSED) {
128 rb_erase(&rp->c_node, &b->rb_head);
129 list_del(&rp->c_lru);
130 atomic_dec(&nn->num_drc_entries);
131 nfsd_stats_drc_mem_usage_sub(nn, sizeof(*rp));
132 }
133 kmem_cache_free(drc_slab, rp);
134}
135
136static void
137nfsd_reply_cache_free(struct nfsd_drc_bucket *b, struct svc_cacherep *rp,
138 struct nfsd_net *nn)
139{
140 spin_lock(&b->cache_lock);
141 nfsd_reply_cache_free_locked(b, rp, nn);
142 spin_unlock(&b->cache_lock);
143}
144
145int nfsd_drc_slab_create(void)
146{
147 drc_slab = kmem_cache_create("nfsd_drc",
148 sizeof(struct svc_cacherep), 0, 0, NULL);
149 return drc_slab ? 0: -ENOMEM;
150}
151
152void nfsd_drc_slab_free(void)
153{
154 kmem_cache_destroy(drc_slab);
155}
156
157static int nfsd_reply_cache_stats_init(struct nfsd_net *nn)
158{
159 return nfsd_percpu_counters_init(nn->counter, NFSD_NET_COUNTERS_NUM);
160}
161
162static void nfsd_reply_cache_stats_destroy(struct nfsd_net *nn)
163{
164 nfsd_percpu_counters_destroy(nn->counter, NFSD_NET_COUNTERS_NUM);
165}
166
167int nfsd_reply_cache_init(struct nfsd_net *nn)
168{
169 unsigned int hashsize;
170 unsigned int i;
171 int status = 0;
172
173 nn->max_drc_entries = nfsd_cache_size_limit();
174 atomic_set(&nn->num_drc_entries, 0);
175 hashsize = nfsd_hashsize(nn->max_drc_entries);
176 nn->maskbits = ilog2(hashsize);
177
178 status = nfsd_reply_cache_stats_init(nn);
179 if (status)
180 goto out_nomem;
181
182 nn->nfsd_reply_cache_shrinker.scan_objects = nfsd_reply_cache_scan;
183 nn->nfsd_reply_cache_shrinker.count_objects = nfsd_reply_cache_count;
184 nn->nfsd_reply_cache_shrinker.seeks = 1;
185 status = register_shrinker(&nn->nfsd_reply_cache_shrinker);
186 if (status)
187 goto out_stats_destroy;
188
189 nn->drc_hashtbl = kvzalloc(array_size(hashsize,
190 sizeof(*nn->drc_hashtbl)), GFP_KERNEL);
191 if (!nn->drc_hashtbl)
192 goto out_shrinker;
193
194 for (i = 0; i < hashsize; i++) {
195 INIT_LIST_HEAD(&nn->drc_hashtbl[i].lru_head);
196 spin_lock_init(&nn->drc_hashtbl[i].cache_lock);
197 }
198 nn->drc_hashsize = hashsize;
199
200 return 0;
201out_shrinker:
202 unregister_shrinker(&nn->nfsd_reply_cache_shrinker);
203out_stats_destroy:
204 nfsd_reply_cache_stats_destroy(nn);
205out_nomem:
206 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
207 return -ENOMEM;
208}
209
210void nfsd_reply_cache_shutdown(struct nfsd_net *nn)
211{
212 struct svc_cacherep *rp;
213 unsigned int i;
214
215 nfsd_reply_cache_stats_destroy(nn);
216 unregister_shrinker(&nn->nfsd_reply_cache_shrinker);
217
218 for (i = 0; i < nn->drc_hashsize; i++) {
219 struct list_head *head = &nn->drc_hashtbl[i].lru_head;
220 while (!list_empty(head)) {
221 rp = list_first_entry(head, struct svc_cacherep, c_lru);
222 nfsd_reply_cache_free_locked(&nn->drc_hashtbl[i],
223 rp, nn);
224 }
225 }
226
227 kvfree(nn->drc_hashtbl);
228 nn->drc_hashtbl = NULL;
229 nn->drc_hashsize = 0;
230
231}
232
233/*
234 * Move cache entry to end of LRU list, and queue the cleaner to run if it's
235 * not already scheduled.
236 */
237static void
238lru_put_end(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
239{
240 rp->c_timestamp = jiffies;
241 list_move_tail(&rp->c_lru, &b->lru_head);
242}
243
244static long
245prune_bucket(struct nfsd_drc_bucket *b, struct nfsd_net *nn)
246{
247 struct svc_cacherep *rp, *tmp;
248 long freed = 0;
249
250 list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
251 /*
252 * Don't free entries attached to calls that are still
253 * in-progress, but do keep scanning the list.
254 */
255 if (rp->c_state == RC_INPROG)
256 continue;
257 if (atomic_read(&nn->num_drc_entries) <= nn->max_drc_entries &&
258 time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
259 break;
260 nfsd_reply_cache_free_locked(b, rp, nn);
261 freed++;
262 }
263 return freed;
264}
265
266/*
267 * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
268 * Also prune the oldest ones when the total exceeds the max number of entries.
269 */
270static long
271prune_cache_entries(struct nfsd_net *nn)
272{
273 unsigned int i;
274 long freed = 0;
275
276 for (i = 0; i < nn->drc_hashsize; i++) {
277 struct nfsd_drc_bucket *b = &nn->drc_hashtbl[i];
278
279 if (list_empty(&b->lru_head))
280 continue;
281 spin_lock(&b->cache_lock);
282 freed += prune_bucket(b, nn);
283 spin_unlock(&b->cache_lock);
284 }
285 return freed;
286}
287
288static unsigned long
289nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
290{
291 struct nfsd_net *nn = container_of(shrink,
292 struct nfsd_net, nfsd_reply_cache_shrinker);
293
294 return atomic_read(&nn->num_drc_entries);
295}
296
297static unsigned long
298nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
299{
300 struct nfsd_net *nn = container_of(shrink,
301 struct nfsd_net, nfsd_reply_cache_shrinker);
302
303 return prune_cache_entries(nn);
304}
305/*
306 * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
307 */
308static __wsum
309nfsd_cache_csum(struct svc_rqst *rqstp)
310{
311 int idx;
312 unsigned int base;
313 __wsum csum;
314 struct xdr_buf *buf = &rqstp->rq_arg;
315 const unsigned char *p = buf->head[0].iov_base;
316 size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
317 RC_CSUMLEN);
318 size_t len = min(buf->head[0].iov_len, csum_len);
319
320 /* rq_arg.head first */
321 csum = csum_partial(p, len, 0);
322 csum_len -= len;
323
324 /* Continue into page array */
325 idx = buf->page_base / PAGE_SIZE;
326 base = buf->page_base & ~PAGE_MASK;
327 while (csum_len) {
328 p = page_address(buf->pages[idx]) + base;
329 len = min_t(size_t, PAGE_SIZE - base, csum_len);
330 csum = csum_partial(p, len, csum);
331 csum_len -= len;
332 base = 0;
333 ++idx;
334 }
335 return csum;
336}
337
338static int
339nfsd_cache_key_cmp(const struct svc_cacherep *key,
340 const struct svc_cacherep *rp, struct nfsd_net *nn)
341{
342 if (key->c_key.k_xid == rp->c_key.k_xid &&
343 key->c_key.k_csum != rp->c_key.k_csum) {
344 nfsd_stats_payload_misses_inc(nn);
345 trace_nfsd_drc_mismatch(nn, key, rp);
346 }
347
348 return memcmp(&key->c_key, &rp->c_key, sizeof(key->c_key));
349}
350
351/*
352 * Search the request hash for an entry that matches the given rqstp.
353 * Must be called with cache_lock held. Returns the found entry or
354 * inserts an empty key on failure.
355 */
356static struct svc_cacherep *
357nfsd_cache_insert(struct nfsd_drc_bucket *b, struct svc_cacherep *key,
358 struct nfsd_net *nn)
359{
360 struct svc_cacherep *rp, *ret = key;
361 struct rb_node **p = &b->rb_head.rb_node,
362 *parent = NULL;
363 unsigned int entries = 0;
364 int cmp;
365
366 while (*p != NULL) {
367 ++entries;
368 parent = *p;
369 rp = rb_entry(parent, struct svc_cacherep, c_node);
370
371 cmp = nfsd_cache_key_cmp(key, rp, nn);
372 if (cmp < 0)
373 p = &parent->rb_left;
374 else if (cmp > 0)
375 p = &parent->rb_right;
376 else {
377 ret = rp;
378 goto out;
379 }
380 }
381 rb_link_node(&key->c_node, parent, p);
382 rb_insert_color(&key->c_node, &b->rb_head);
383out:
384 /* tally hash chain length stats */
385 if (entries > nn->longest_chain) {
386 nn->longest_chain = entries;
387 nn->longest_chain_cachesize = atomic_read(&nn->num_drc_entries);
388 } else if (entries == nn->longest_chain) {
389 /* prefer to keep the smallest cachesize possible here */
390 nn->longest_chain_cachesize = min_t(unsigned int,
391 nn->longest_chain_cachesize,
392 atomic_read(&nn->num_drc_entries));
393 }
394
395 lru_put_end(b, ret);
396 return ret;
397}
398
399/**
400 * nfsd_cache_lookup - Find an entry in the duplicate reply cache
401 * @rqstp: Incoming Call to find
402 *
403 * Try to find an entry matching the current call in the cache. When none
404 * is found, we try to grab the oldest expired entry off the LRU list. If
405 * a suitable one isn't there, then drop the cache_lock and allocate a
406 * new one, then search again in case one got inserted while this thread
407 * didn't hold the lock.
408 *
409 * Return values:
410 * %RC_DOIT: Process the request normally
411 * %RC_REPLY: Reply from cache
412 * %RC_DROPIT: Do not process the request further
413 */
414int nfsd_cache_lookup(struct svc_rqst *rqstp)
415{
416 struct nfsd_net *nn = net_generic(SVC_NET(rqstp), nfsd_net_id);
417 struct svc_cacherep *rp, *found;
418 __be32 xid = rqstp->rq_xid;
419 __wsum csum;
420 u32 hash = nfsd_cache_hash(xid, nn);
421 struct nfsd_drc_bucket *b = &nn->drc_hashtbl[hash];
422 int type = rqstp->rq_cachetype;
423 int rtn = RC_DOIT;
424
425 rqstp->rq_cacherep = NULL;
426 if (type == RC_NOCACHE) {
427 nfsd_stats_rc_nocache_inc();
428 goto out;
429 }
430
431 csum = nfsd_cache_csum(rqstp);
432
433 /*
434 * Since the common case is a cache miss followed by an insert,
435 * preallocate an entry.
436 */
437 rp = nfsd_reply_cache_alloc(rqstp, csum, nn);
438 if (!rp)
439 goto out;
440
441 spin_lock(&b->cache_lock);
442 found = nfsd_cache_insert(b, rp, nn);
443 if (found != rp) {
444 nfsd_reply_cache_free_locked(NULL, rp, nn);
445 rp = found;
446 goto found_entry;
447 }
448
449 nfsd_stats_rc_misses_inc();
450 rqstp->rq_cacherep = rp;
451 rp->c_state = RC_INPROG;
452
453 atomic_inc(&nn->num_drc_entries);
454 nfsd_stats_drc_mem_usage_add(nn, sizeof(*rp));
455
456 /* go ahead and prune the cache */
457 prune_bucket(b, nn);
458
459out_unlock:
460 spin_unlock(&b->cache_lock);
461out:
462 return rtn;
463
464found_entry:
465 /* We found a matching entry which is either in progress or done. */
466 nfsd_stats_rc_hits_inc();
467 rtn = RC_DROPIT;
468
469 /* Request being processed */
470 if (rp->c_state == RC_INPROG)
471 goto out_trace;
472
473 /* From the hall of fame of impractical attacks:
474 * Is this a user who tries to snoop on the cache? */
475 rtn = RC_DOIT;
476 if (!test_bit(RQ_SECURE, &rqstp->rq_flags) && rp->c_secure)
477 goto out_trace;
478
479 /* Compose RPC reply header */
480 switch (rp->c_type) {
481 case RC_NOCACHE:
482 break;
483 case RC_REPLSTAT:
484 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
485 rtn = RC_REPLY;
486 break;
487 case RC_REPLBUFF:
488 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
489 goto out_unlock; /* should not happen */
490 rtn = RC_REPLY;
491 break;
492 default:
493 WARN_ONCE(1, "nfsd: bad repcache type %d\n", rp->c_type);
494 }
495
496out_trace:
497 trace_nfsd_drc_found(nn, rqstp, rtn);
498 goto out_unlock;
499}
500
501/**
502 * nfsd_cache_update - Update an entry in the duplicate reply cache.
503 * @rqstp: svc_rqst with a finished Reply
504 * @cachetype: which cache to update
505 * @statp: Reply's status code
506 *
507 * This is called from nfsd_dispatch when the procedure has been
508 * executed and the complete reply is in rqstp->rq_res.
509 *
510 * We're copying around data here rather than swapping buffers because
511 * the toplevel loop requires max-sized buffers, which would be a waste
512 * of memory for a cache with a max reply size of 100 bytes (diropokres).
513 *
514 * If we should start to use different types of cache entries tailored
515 * specifically for attrstat and fh's, we may save even more space.
516 *
517 * Also note that a cachetype of RC_NOCACHE can legally be passed when
518 * nfsd failed to encode a reply that otherwise would have been cached.
519 * In this case, nfsd_cache_update is called with statp == NULL.
520 */
521void nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
522{
523 struct nfsd_net *nn = net_generic(SVC_NET(rqstp), nfsd_net_id);
524 struct svc_cacherep *rp = rqstp->rq_cacherep;
525 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
526 u32 hash;
527 struct nfsd_drc_bucket *b;
528 int len;
529 size_t bufsize = 0;
530
531 if (!rp)
532 return;
533
534 hash = nfsd_cache_hash(rp->c_key.k_xid, nn);
535 b = &nn->drc_hashtbl[hash];
536
537 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
538 len >>= 2;
539
540 /* Don't cache excessive amounts of data and XDR failures */
541 if (!statp || len > (256 >> 2)) {
542 nfsd_reply_cache_free(b, rp, nn);
543 return;
544 }
545
546 switch (cachetype) {
547 case RC_REPLSTAT:
548 if (len != 1)
549 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
550 rp->c_replstat = *statp;
551 break;
552 case RC_REPLBUFF:
553 cachv = &rp->c_replvec;
554 bufsize = len << 2;
555 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
556 if (!cachv->iov_base) {
557 nfsd_reply_cache_free(b, rp, nn);
558 return;
559 }
560 cachv->iov_len = bufsize;
561 memcpy(cachv->iov_base, statp, bufsize);
562 break;
563 case RC_NOCACHE:
564 nfsd_reply_cache_free(b, rp, nn);
565 return;
566 }
567 spin_lock(&b->cache_lock);
568 nfsd_stats_drc_mem_usage_add(nn, bufsize);
569 lru_put_end(b, rp);
570 rp->c_secure = test_bit(RQ_SECURE, &rqstp->rq_flags);
571 rp->c_type = cachetype;
572 rp->c_state = RC_DONE;
573 spin_unlock(&b->cache_lock);
574 return;
575}
576
577/*
578 * Copy cached reply to current reply buffer. Should always fit.
579 * FIXME as reply is in a page, we should just attach the page, and
580 * keep a refcount....
581 */
582static int
583nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
584{
585 struct kvec *vec = &rqstp->rq_res.head[0];
586
587 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
588 printk(KERN_WARNING "nfsd: cached reply too large (%zd).\n",
589 data->iov_len);
590 return 0;
591 }
592 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
593 vec->iov_len += data->iov_len;
594 return 1;
595}
596
597/*
598 * Note that fields may be added, removed or reordered in the future. Programs
599 * scraping this file for info should test the labels to ensure they're
600 * getting the correct field.
601 */
602static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
603{
604 struct nfsd_net *nn = m->private;
605
606 seq_printf(m, "max entries: %u\n", nn->max_drc_entries);
607 seq_printf(m, "num entries: %u\n",
608 atomic_read(&nn->num_drc_entries));
609 seq_printf(m, "hash buckets: %u\n", 1 << nn->maskbits);
610 seq_printf(m, "mem usage: %lld\n",
611 percpu_counter_sum_positive(&nn->counter[NFSD_NET_DRC_MEM_USAGE]));
612 seq_printf(m, "cache hits: %lld\n",
613 percpu_counter_sum_positive(&nfsdstats.counter[NFSD_STATS_RC_HITS]));
614 seq_printf(m, "cache misses: %lld\n",
615 percpu_counter_sum_positive(&nfsdstats.counter[NFSD_STATS_RC_MISSES]));
616 seq_printf(m, "not cached: %lld\n",
617 percpu_counter_sum_positive(&nfsdstats.counter[NFSD_STATS_RC_NOCACHE]));
618 seq_printf(m, "payload misses: %lld\n",
619 percpu_counter_sum_positive(&nn->counter[NFSD_NET_PAYLOAD_MISSES]));
620 seq_printf(m, "longest chain len: %u\n", nn->longest_chain);
621 seq_printf(m, "cachesize at longest: %u\n", nn->longest_chain_cachesize);
622 return 0;
623}
624
625int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
626{
627 struct nfsd_net *nn = net_generic(file_inode(file)->i_sb->s_fs_info,
628 nfsd_net_id);
629
630 return single_open(file, nfsd_reply_cache_stats_show, nn);
631}