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
30struct nfsd_drc_bucket {
31 struct list_head lru_head;
32 spinlock_t cache_lock;
33};
34
35static struct nfsd_drc_bucket *drc_hashtbl;
36static struct kmem_cache *drc_slab;
37
38/* max number of entries allowed in the cache */
39static unsigned int max_drc_entries;
40
41/* number of significant bits in the hash value */
42static unsigned int maskbits;
43static unsigned int drc_hashsize;
44
45/*
46 * Stats and other tracking of on the duplicate reply cache. All of these and
47 * the "rc" fields in nfsdstats are protected by the cache_lock
48 */
49
50/* total number of entries */
51static atomic_t num_drc_entries;
52
53/* cache misses due only to checksum comparison failures */
54static unsigned int payload_misses;
55
56/* amount of memory (in bytes) currently consumed by the DRC */
57static unsigned int drc_mem_usage;
58
59/* longest hash chain seen */
60static unsigned int longest_chain;
61
62/* size of cache when we saw the longest hash chain */
63static unsigned int longest_chain_cachesize;
64
65static int nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
66static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
67 struct shrink_control *sc);
68static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
69 struct shrink_control *sc);
70
71static struct shrinker nfsd_reply_cache_shrinker = {
72 .scan_objects = nfsd_reply_cache_scan,
73 .count_objects = nfsd_reply_cache_count,
74 .seeks = 1,
75};
76
77/*
78 * Put a cap on the size of the DRC based on the amount of available
79 * low memory in the machine.
80 *
81 * 64MB: 8192
82 * 128MB: 11585
83 * 256MB: 16384
84 * 512MB: 23170
85 * 1GB: 32768
86 * 2GB: 46340
87 * 4GB: 65536
88 * 8GB: 92681
89 * 16GB: 131072
90 *
91 * ...with a hard cap of 256k entries. In the worst case, each entry will be
92 * ~1k, so the above numbers should give a rough max of the amount of memory
93 * used in k.
94 */
95static unsigned int
96nfsd_cache_size_limit(void)
97{
98 unsigned int limit;
99 unsigned long low_pages = totalram_pages - totalhigh_pages;
100
101 limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
102 return min_t(unsigned int, limit, 256*1024);
103}
104
105/*
106 * Compute the number of hash buckets we need. Divide the max cachesize by
107 * the "target" max bucket size, and round up to next power of two.
108 */
109static unsigned int
110nfsd_hashsize(unsigned int limit)
111{
112 return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
113}
114
115static u32
116nfsd_cache_hash(__be32 xid)
117{
118 return hash_32(be32_to_cpu(xid), maskbits);
119}
120
121static struct svc_cacherep *
122nfsd_reply_cache_alloc(void)
123{
124 struct svc_cacherep *rp;
125
126 rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
127 if (rp) {
128 rp->c_state = RC_UNUSED;
129 rp->c_type = RC_NOCACHE;
130 INIT_LIST_HEAD(&rp->c_lru);
131 }
132 return rp;
133}
134
135static void
136nfsd_reply_cache_free_locked(struct svc_cacherep *rp)
137{
138 if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
139 drc_mem_usage -= rp->c_replvec.iov_len;
140 kfree(rp->c_replvec.iov_base);
141 }
142 list_del(&rp->c_lru);
143 atomic_dec(&num_drc_entries);
144 drc_mem_usage -= sizeof(*rp);
145 kmem_cache_free(drc_slab, rp);
146}
147
148static void
149nfsd_reply_cache_free(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
150{
151 spin_lock(&b->cache_lock);
152 nfsd_reply_cache_free_locked(rp);
153 spin_unlock(&b->cache_lock);
154}
155
156int nfsd_reply_cache_init(void)
157{
158 unsigned int hashsize;
159 unsigned int i;
160 int status = 0;
161
162 max_drc_entries = nfsd_cache_size_limit();
163 atomic_set(&num_drc_entries, 0);
164 hashsize = nfsd_hashsize(max_drc_entries);
165 maskbits = ilog2(hashsize);
166
167 status = register_shrinker(&nfsd_reply_cache_shrinker);
168 if (status)
169 return status;
170
171 drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
172 0, 0, NULL);
173 if (!drc_slab)
174 goto out_nomem;
175
176 drc_hashtbl = kcalloc(hashsize, sizeof(*drc_hashtbl), GFP_KERNEL);
177 if (!drc_hashtbl)
178 goto out_nomem;
179 for (i = 0; i < hashsize; i++) {
180 INIT_LIST_HEAD(&drc_hashtbl[i].lru_head);
181 spin_lock_init(&drc_hashtbl[i].cache_lock);
182 }
183 drc_hashsize = hashsize;
184
185 return 0;
186out_nomem:
187 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
188 nfsd_reply_cache_shutdown();
189 return -ENOMEM;
190}
191
192void nfsd_reply_cache_shutdown(void)
193{
194 struct svc_cacherep *rp;
195 unsigned int i;
196
197 unregister_shrinker(&nfsd_reply_cache_shrinker);
198
199 for (i = 0; i < drc_hashsize; i++) {
200 struct list_head *head = &drc_hashtbl[i].lru_head;
201 while (!list_empty(head)) {
202 rp = list_first_entry(head, struct svc_cacherep, c_lru);
203 nfsd_reply_cache_free_locked(rp);
204 }
205 }
206
207 kfree (drc_hashtbl);
208 drc_hashtbl = NULL;
209 drc_hashsize = 0;
210
211 kmem_cache_destroy(drc_slab);
212 drc_slab = NULL;
213}
214
215/*
216 * Move cache entry to end of LRU list, and queue the cleaner to run if it's
217 * not already scheduled.
218 */
219static void
220lru_put_end(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
221{
222 rp->c_timestamp = jiffies;
223 list_move_tail(&rp->c_lru, &b->lru_head);
224}
225
226static long
227prune_bucket(struct nfsd_drc_bucket *b)
228{
229 struct svc_cacherep *rp, *tmp;
230 long freed = 0;
231
232 list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
233 /*
234 * Don't free entries attached to calls that are still
235 * in-progress, but do keep scanning the list.
236 */
237 if (rp->c_state == RC_INPROG)
238 continue;
239 if (atomic_read(&num_drc_entries) <= max_drc_entries &&
240 time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
241 break;
242 nfsd_reply_cache_free_locked(rp);
243 freed++;
244 }
245 return freed;
246}
247
248/*
249 * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
250 * Also prune the oldest ones when the total exceeds the max number of entries.
251 */
252static long
253prune_cache_entries(void)
254{
255 unsigned int i;
256 long freed = 0;
257
258 for (i = 0; i < drc_hashsize; i++) {
259 struct nfsd_drc_bucket *b = &drc_hashtbl[i];
260
261 if (list_empty(&b->lru_head))
262 continue;
263 spin_lock(&b->cache_lock);
264 freed += prune_bucket(b);
265 spin_unlock(&b->cache_lock);
266 }
267 return freed;
268}
269
270static unsigned long
271nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
272{
273 return atomic_read(&num_drc_entries);
274}
275
276static unsigned long
277nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
278{
279 return prune_cache_entries();
280}
281/*
282 * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
283 */
284static __wsum
285nfsd_cache_csum(struct svc_rqst *rqstp)
286{
287 int idx;
288 unsigned int base;
289 __wsum csum;
290 struct xdr_buf *buf = &rqstp->rq_arg;
291 const unsigned char *p = buf->head[0].iov_base;
292 size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
293 RC_CSUMLEN);
294 size_t len = min(buf->head[0].iov_len, csum_len);
295
296 /* rq_arg.head first */
297 csum = csum_partial(p, len, 0);
298 csum_len -= len;
299
300 /* Continue into page array */
301 idx = buf->page_base / PAGE_SIZE;
302 base = buf->page_base & ~PAGE_MASK;
303 while (csum_len) {
304 p = page_address(buf->pages[idx]) + base;
305 len = min_t(size_t, PAGE_SIZE - base, csum_len);
306 csum = csum_partial(p, len, csum);
307 csum_len -= len;
308 base = 0;
309 ++idx;
310 }
311 return csum;
312}
313
314static bool
315nfsd_cache_match(struct svc_rqst *rqstp, __wsum csum, struct svc_cacherep *rp)
316{
317 /* Check RPC XID first */
318 if (rqstp->rq_xid != rp->c_xid)
319 return false;
320 /* compare checksum of NFS data */
321 if (csum != rp->c_csum) {
322 ++payload_misses;
323 return false;
324 }
325
326 /* Other discriminators */
327 if (rqstp->rq_proc != rp->c_proc ||
328 rqstp->rq_prot != rp->c_prot ||
329 rqstp->rq_vers != rp->c_vers ||
330 rqstp->rq_arg.len != rp->c_len ||
331 !rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) ||
332 rpc_get_port(svc_addr(rqstp)) != rpc_get_port((struct sockaddr *)&rp->c_addr))
333 return false;
334
335 return true;
336}
337
338/*
339 * Search the request hash for an entry that matches the given rqstp.
340 * Must be called with cache_lock held. Returns the found entry or
341 * NULL on failure.
342 */
343static struct svc_cacherep *
344nfsd_cache_search(struct nfsd_drc_bucket *b, struct svc_rqst *rqstp,
345 __wsum csum)
346{
347 struct svc_cacherep *rp, *ret = NULL;
348 struct list_head *rh = &b->lru_head;
349 unsigned int entries = 0;
350
351 list_for_each_entry(rp, rh, c_lru) {
352 ++entries;
353 if (nfsd_cache_match(rqstp, csum, rp)) {
354 ret = rp;
355 break;
356 }
357 }
358
359 /* tally hash chain length stats */
360 if (entries > longest_chain) {
361 longest_chain = entries;
362 longest_chain_cachesize = atomic_read(&num_drc_entries);
363 } else if (entries == longest_chain) {
364 /* prefer to keep the smallest cachesize possible here */
365 longest_chain_cachesize = min_t(unsigned int,
366 longest_chain_cachesize,
367 atomic_read(&num_drc_entries));
368 }
369
370 return ret;
371}
372
373/*
374 * Try to find an entry matching the current call in the cache. When none
375 * is found, we try to grab the oldest expired entry off the LRU list. If
376 * a suitable one isn't there, then drop the cache_lock and allocate a
377 * new one, then search again in case one got inserted while this thread
378 * didn't hold the lock.
379 */
380int
381nfsd_cache_lookup(struct svc_rqst *rqstp)
382{
383 struct svc_cacherep *rp, *found;
384 __be32 xid = rqstp->rq_xid;
385 u32 proto = rqstp->rq_prot,
386 vers = rqstp->rq_vers,
387 proc = rqstp->rq_proc;
388 __wsum csum;
389 u32 hash = nfsd_cache_hash(xid);
390 struct nfsd_drc_bucket *b = &drc_hashtbl[hash];
391 unsigned long age;
392 int type = rqstp->rq_cachetype;
393 int rtn = RC_DOIT;
394
395 rqstp->rq_cacherep = NULL;
396 if (type == RC_NOCACHE) {
397 nfsdstats.rcnocache++;
398 return rtn;
399 }
400
401 csum = nfsd_cache_csum(rqstp);
402
403 /*
404 * Since the common case is a cache miss followed by an insert,
405 * preallocate an entry.
406 */
407 rp = nfsd_reply_cache_alloc();
408 spin_lock(&b->cache_lock);
409 if (likely(rp)) {
410 atomic_inc(&num_drc_entries);
411 drc_mem_usage += sizeof(*rp);
412 }
413
414 /* go ahead and prune the cache */
415 prune_bucket(b);
416
417 found = nfsd_cache_search(b, rqstp, csum);
418 if (found) {
419 if (likely(rp))
420 nfsd_reply_cache_free_locked(rp);
421 rp = found;
422 goto found_entry;
423 }
424
425 if (!rp) {
426 dprintk("nfsd: unable to allocate DRC entry!\n");
427 goto out;
428 }
429
430 nfsdstats.rcmisses++;
431 rqstp->rq_cacherep = rp;
432 rp->c_state = RC_INPROG;
433 rp->c_xid = xid;
434 rp->c_proc = proc;
435 rpc_copy_addr((struct sockaddr *)&rp->c_addr, svc_addr(rqstp));
436 rpc_set_port((struct sockaddr *)&rp->c_addr, rpc_get_port(svc_addr(rqstp)));
437 rp->c_prot = proto;
438 rp->c_vers = vers;
439 rp->c_len = rqstp->rq_arg.len;
440 rp->c_csum = csum;
441
442 lru_put_end(b, rp);
443
444 /* release any buffer */
445 if (rp->c_type == RC_REPLBUFF) {
446 drc_mem_usage -= rp->c_replvec.iov_len;
447 kfree(rp->c_replvec.iov_base);
448 rp->c_replvec.iov_base = NULL;
449 }
450 rp->c_type = RC_NOCACHE;
451 out:
452 spin_unlock(&b->cache_lock);
453 return rtn;
454
455found_entry:
456 nfsdstats.rchits++;
457 /* We found a matching entry which is either in progress or done. */
458 age = jiffies - rp->c_timestamp;
459 lru_put_end(b, rp);
460
461 rtn = RC_DROPIT;
462 /* Request being processed or excessive rexmits */
463 if (rp->c_state == RC_INPROG || age < RC_DELAY)
464 goto out;
465
466 /* From the hall of fame of impractical attacks:
467 * Is this a user who tries to snoop on the cache? */
468 rtn = RC_DOIT;
469 if (!test_bit(RQ_SECURE, &rqstp->rq_flags) && rp->c_secure)
470 goto out;
471
472 /* Compose RPC reply header */
473 switch (rp->c_type) {
474 case RC_NOCACHE:
475 break;
476 case RC_REPLSTAT:
477 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
478 rtn = RC_REPLY;
479 break;
480 case RC_REPLBUFF:
481 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
482 goto out; /* should not happen */
483 rtn = RC_REPLY;
484 break;
485 default:
486 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
487 nfsd_reply_cache_free_locked(rp);
488 }
489
490 goto out;
491}
492
493/*
494 * Update a cache entry. This is called from nfsd_dispatch when
495 * the procedure has been executed and the complete reply is in
496 * rqstp->rq_res.
497 *
498 * We're copying around data here rather than swapping buffers because
499 * the toplevel loop requires max-sized buffers, which would be a waste
500 * of memory for a cache with a max reply size of 100 bytes (diropokres).
501 *
502 * If we should start to use different types of cache entries tailored
503 * specifically for attrstat and fh's, we may save even more space.
504 *
505 * Also note that a cachetype of RC_NOCACHE can legally be passed when
506 * nfsd failed to encode a reply that otherwise would have been cached.
507 * In this case, nfsd_cache_update is called with statp == NULL.
508 */
509void
510nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
511{
512 struct svc_cacherep *rp = rqstp->rq_cacherep;
513 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
514 u32 hash;
515 struct nfsd_drc_bucket *b;
516 int len;
517 size_t bufsize = 0;
518
519 if (!rp)
520 return;
521
522 hash = nfsd_cache_hash(rp->c_xid);
523 b = &drc_hashtbl[hash];
524
525 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
526 len >>= 2;
527
528 /* Don't cache excessive amounts of data and XDR failures */
529 if (!statp || len > (256 >> 2)) {
530 nfsd_reply_cache_free(b, rp);
531 return;
532 }
533
534 switch (cachetype) {
535 case RC_REPLSTAT:
536 if (len != 1)
537 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
538 rp->c_replstat = *statp;
539 break;
540 case RC_REPLBUFF:
541 cachv = &rp->c_replvec;
542 bufsize = len << 2;
543 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
544 if (!cachv->iov_base) {
545 nfsd_reply_cache_free(b, rp);
546 return;
547 }
548 cachv->iov_len = bufsize;
549 memcpy(cachv->iov_base, statp, bufsize);
550 break;
551 case RC_NOCACHE:
552 nfsd_reply_cache_free(b, rp);
553 return;
554 }
555 spin_lock(&b->cache_lock);
556 drc_mem_usage += bufsize;
557 lru_put_end(b, rp);
558 rp->c_secure = test_bit(RQ_SECURE, &rqstp->rq_flags);
559 rp->c_type = cachetype;
560 rp->c_state = RC_DONE;
561 spin_unlock(&b->cache_lock);
562 return;
563}
564
565/*
566 * Copy cached reply to current reply buffer. Should always fit.
567 * FIXME as reply is in a page, we should just attach the page, and
568 * keep a refcount....
569 */
570static int
571nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
572{
573 struct kvec *vec = &rqstp->rq_res.head[0];
574
575 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
576 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
577 data->iov_len);
578 return 0;
579 }
580 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
581 vec->iov_len += data->iov_len;
582 return 1;
583}
584
585/*
586 * Note that fields may be added, removed or reordered in the future. Programs
587 * scraping this file for info should test the labels to ensure they're
588 * getting the correct field.
589 */
590static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
591{
592 seq_printf(m, "max entries: %u\n", max_drc_entries);
593 seq_printf(m, "num entries: %u\n",
594 atomic_read(&num_drc_entries));
595 seq_printf(m, "hash buckets: %u\n", 1 << maskbits);
596 seq_printf(m, "mem usage: %u\n", drc_mem_usage);
597 seq_printf(m, "cache hits: %u\n", nfsdstats.rchits);
598 seq_printf(m, "cache misses: %u\n", nfsdstats.rcmisses);
599 seq_printf(m, "not cached: %u\n", nfsdstats.rcnocache);
600 seq_printf(m, "payload misses: %u\n", payload_misses);
601 seq_printf(m, "longest chain len: %u\n", longest_chain);
602 seq_printf(m, "cachesize at longest: %u\n", longest_chain_cachesize);
603 return 0;
604}
605
606int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
607{
608 return single_open(file, nfsd_reply_cache_stats_show, NULL);
609}
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/vmalloc.h>
13#include <linux/sunrpc/addr.h>
14#include <linux/highmem.h>
15#include <linux/log2.h>
16#include <linux/hash.h>
17#include <net/checksum.h>
18
19#include "nfsd.h"
20#include "cache.h"
21
22#define NFSDDBG_FACILITY NFSDDBG_REPCACHE
23
24/*
25 * We use this value to determine the number of hash buckets from the max
26 * cache size, the idea being that when the cache is at its maximum number
27 * of entries, then this should be the average number of entries per bucket.
28 */
29#define TARGET_BUCKET_SIZE 64
30
31struct nfsd_drc_bucket {
32 struct list_head lru_head;
33 spinlock_t cache_lock;
34};
35
36static struct nfsd_drc_bucket *drc_hashtbl;
37static struct kmem_cache *drc_slab;
38
39/* max number of entries allowed in the cache */
40static unsigned int max_drc_entries;
41
42/* number of significant bits in the hash value */
43static unsigned int maskbits;
44static unsigned int drc_hashsize;
45
46/*
47 * Stats and other tracking of on the duplicate reply cache. All of these and
48 * the "rc" fields in nfsdstats are protected by the cache_lock
49 */
50
51/* total number of entries */
52static atomic_t num_drc_entries;
53
54/* cache misses due only to checksum comparison failures */
55static unsigned int payload_misses;
56
57/* amount of memory (in bytes) currently consumed by the DRC */
58static unsigned int drc_mem_usage;
59
60/* longest hash chain seen */
61static unsigned int longest_chain;
62
63/* size of cache when we saw the longest hash chain */
64static unsigned int longest_chain_cachesize;
65
66static int nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
67static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
68 struct shrink_control *sc);
69static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
70 struct shrink_control *sc);
71
72static struct shrinker nfsd_reply_cache_shrinker = {
73 .scan_objects = nfsd_reply_cache_scan,
74 .count_objects = nfsd_reply_cache_count,
75 .seeks = 1,
76};
77
78/*
79 * Put a cap on the size of the DRC based on the amount of available
80 * low memory in the machine.
81 *
82 * 64MB: 8192
83 * 128MB: 11585
84 * 256MB: 16384
85 * 512MB: 23170
86 * 1GB: 32768
87 * 2GB: 46340
88 * 4GB: 65536
89 * 8GB: 92681
90 * 16GB: 131072
91 *
92 * ...with a hard cap of 256k entries. In the worst case, each entry will be
93 * ~1k, so the above numbers should give a rough max of the amount of memory
94 * used in k.
95 */
96static unsigned int
97nfsd_cache_size_limit(void)
98{
99 unsigned int limit;
100 unsigned long low_pages = totalram_pages - totalhigh_pages;
101
102 limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
103 return min_t(unsigned int, limit, 256*1024);
104}
105
106/*
107 * Compute the number of hash buckets we need. Divide the max cachesize by
108 * the "target" max bucket size, and round up to next power of two.
109 */
110static unsigned int
111nfsd_hashsize(unsigned int limit)
112{
113 return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
114}
115
116static u32
117nfsd_cache_hash(__be32 xid)
118{
119 return hash_32(be32_to_cpu(xid), maskbits);
120}
121
122static struct svc_cacherep *
123nfsd_reply_cache_alloc(void)
124{
125 struct svc_cacherep *rp;
126
127 rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
128 if (rp) {
129 rp->c_state = RC_UNUSED;
130 rp->c_type = RC_NOCACHE;
131 INIT_LIST_HEAD(&rp->c_lru);
132 }
133 return rp;
134}
135
136static void
137nfsd_reply_cache_free_locked(struct svc_cacherep *rp)
138{
139 if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
140 drc_mem_usage -= rp->c_replvec.iov_len;
141 kfree(rp->c_replvec.iov_base);
142 }
143 list_del(&rp->c_lru);
144 atomic_dec(&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 nfsd_drc_bucket *b, struct svc_cacherep *rp)
151{
152 spin_lock(&b->cache_lock);
153 nfsd_reply_cache_free_locked(rp);
154 spin_unlock(&b->cache_lock);
155}
156
157int nfsd_reply_cache_init(void)
158{
159 unsigned int hashsize;
160 unsigned int i;
161 int status = 0;
162
163 max_drc_entries = nfsd_cache_size_limit();
164 atomic_set(&num_drc_entries, 0);
165 hashsize = nfsd_hashsize(max_drc_entries);
166 maskbits = ilog2(hashsize);
167
168 status = register_shrinker(&nfsd_reply_cache_shrinker);
169 if (status)
170 return status;
171
172 drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
173 0, 0, NULL);
174 if (!drc_slab)
175 goto out_nomem;
176
177 drc_hashtbl = kcalloc(hashsize, sizeof(*drc_hashtbl), GFP_KERNEL);
178 if (!drc_hashtbl) {
179 drc_hashtbl = vzalloc(hashsize * sizeof(*drc_hashtbl));
180 if (!drc_hashtbl)
181 goto out_nomem;
182 }
183
184 for (i = 0; i < hashsize; i++) {
185 INIT_LIST_HEAD(&drc_hashtbl[i].lru_head);
186 spin_lock_init(&drc_hashtbl[i].cache_lock);
187 }
188 drc_hashsize = hashsize;
189
190 return 0;
191out_nomem:
192 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
193 nfsd_reply_cache_shutdown();
194 return -ENOMEM;
195}
196
197void nfsd_reply_cache_shutdown(void)
198{
199 struct svc_cacherep *rp;
200 unsigned int i;
201
202 unregister_shrinker(&nfsd_reply_cache_shrinker);
203
204 for (i = 0; i < drc_hashsize; i++) {
205 struct list_head *head = &drc_hashtbl[i].lru_head;
206 while (!list_empty(head)) {
207 rp = list_first_entry(head, struct svc_cacherep, c_lru);
208 nfsd_reply_cache_free_locked(rp);
209 }
210 }
211
212 kvfree(drc_hashtbl);
213 drc_hashtbl = NULL;
214 drc_hashsize = 0;
215
216 kmem_cache_destroy(drc_slab);
217 drc_slab = NULL;
218}
219
220/*
221 * Move cache entry to end of LRU list, and queue the cleaner to run if it's
222 * not already scheduled.
223 */
224static void
225lru_put_end(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
226{
227 rp->c_timestamp = jiffies;
228 list_move_tail(&rp->c_lru, &b->lru_head);
229}
230
231static long
232prune_bucket(struct nfsd_drc_bucket *b)
233{
234 struct svc_cacherep *rp, *tmp;
235 long freed = 0;
236
237 list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
238 /*
239 * Don't free entries attached to calls that are still
240 * in-progress, but do keep scanning the list.
241 */
242 if (rp->c_state == RC_INPROG)
243 continue;
244 if (atomic_read(&num_drc_entries) <= max_drc_entries &&
245 time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
246 break;
247 nfsd_reply_cache_free_locked(rp);
248 freed++;
249 }
250 return freed;
251}
252
253/*
254 * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
255 * Also prune the oldest ones when the total exceeds the max number of entries.
256 */
257static long
258prune_cache_entries(void)
259{
260 unsigned int i;
261 long freed = 0;
262
263 for (i = 0; i < drc_hashsize; i++) {
264 struct nfsd_drc_bucket *b = &drc_hashtbl[i];
265
266 if (list_empty(&b->lru_head))
267 continue;
268 spin_lock(&b->cache_lock);
269 freed += prune_bucket(b);
270 spin_unlock(&b->cache_lock);
271 }
272 return freed;
273}
274
275static unsigned long
276nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
277{
278 return atomic_read(&num_drc_entries);
279}
280
281static unsigned long
282nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
283{
284 return prune_cache_entries();
285}
286/*
287 * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
288 */
289static __wsum
290nfsd_cache_csum(struct svc_rqst *rqstp)
291{
292 int idx;
293 unsigned int base;
294 __wsum csum;
295 struct xdr_buf *buf = &rqstp->rq_arg;
296 const unsigned char *p = buf->head[0].iov_base;
297 size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
298 RC_CSUMLEN);
299 size_t len = min(buf->head[0].iov_len, csum_len);
300
301 /* rq_arg.head first */
302 csum = csum_partial(p, len, 0);
303 csum_len -= len;
304
305 /* Continue into page array */
306 idx = buf->page_base / PAGE_SIZE;
307 base = buf->page_base & ~PAGE_MASK;
308 while (csum_len) {
309 p = page_address(buf->pages[idx]) + base;
310 len = min_t(size_t, PAGE_SIZE - base, csum_len);
311 csum = csum_partial(p, len, csum);
312 csum_len -= len;
313 base = 0;
314 ++idx;
315 }
316 return csum;
317}
318
319static bool
320nfsd_cache_match(struct svc_rqst *rqstp, __wsum csum, struct svc_cacherep *rp)
321{
322 /* Check RPC XID first */
323 if (rqstp->rq_xid != rp->c_xid)
324 return false;
325 /* compare checksum of NFS data */
326 if (csum != rp->c_csum) {
327 ++payload_misses;
328 return false;
329 }
330
331 /* Other discriminators */
332 if (rqstp->rq_proc != rp->c_proc ||
333 rqstp->rq_prot != rp->c_prot ||
334 rqstp->rq_vers != rp->c_vers ||
335 rqstp->rq_arg.len != rp->c_len ||
336 !rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) ||
337 rpc_get_port(svc_addr(rqstp)) != rpc_get_port((struct sockaddr *)&rp->c_addr))
338 return false;
339
340 return true;
341}
342
343/*
344 * Search the request hash for an entry that matches the given rqstp.
345 * Must be called with cache_lock held. Returns the found entry or
346 * NULL on failure.
347 */
348static struct svc_cacherep *
349nfsd_cache_search(struct nfsd_drc_bucket *b, struct svc_rqst *rqstp,
350 __wsum csum)
351{
352 struct svc_cacherep *rp, *ret = NULL;
353 struct list_head *rh = &b->lru_head;
354 unsigned int entries = 0;
355
356 list_for_each_entry(rp, rh, c_lru) {
357 ++entries;
358 if (nfsd_cache_match(rqstp, csum, rp)) {
359 ret = rp;
360 break;
361 }
362 }
363
364 /* tally hash chain length stats */
365 if (entries > longest_chain) {
366 longest_chain = entries;
367 longest_chain_cachesize = atomic_read(&num_drc_entries);
368 } else if (entries == longest_chain) {
369 /* prefer to keep the smallest cachesize possible here */
370 longest_chain_cachesize = min_t(unsigned int,
371 longest_chain_cachesize,
372 atomic_read(&num_drc_entries));
373 }
374
375 return ret;
376}
377
378/*
379 * Try to find an entry matching the current call in the cache. When none
380 * is found, we try to grab the oldest expired entry off the LRU list. If
381 * a suitable one isn't there, then drop the cache_lock and allocate a
382 * new one, then search again in case one got inserted while this thread
383 * didn't hold the lock.
384 */
385int
386nfsd_cache_lookup(struct svc_rqst *rqstp)
387{
388 struct svc_cacherep *rp, *found;
389 __be32 xid = rqstp->rq_xid;
390 u32 proto = rqstp->rq_prot,
391 vers = rqstp->rq_vers,
392 proc = rqstp->rq_proc;
393 __wsum csum;
394 u32 hash = nfsd_cache_hash(xid);
395 struct nfsd_drc_bucket *b = &drc_hashtbl[hash];
396 unsigned long age;
397 int type = rqstp->rq_cachetype;
398 int rtn = RC_DOIT;
399
400 rqstp->rq_cacherep = NULL;
401 if (type == RC_NOCACHE) {
402 nfsdstats.rcnocache++;
403 return rtn;
404 }
405
406 csum = nfsd_cache_csum(rqstp);
407
408 /*
409 * Since the common case is a cache miss followed by an insert,
410 * preallocate an entry.
411 */
412 rp = nfsd_reply_cache_alloc();
413 spin_lock(&b->cache_lock);
414 if (likely(rp)) {
415 atomic_inc(&num_drc_entries);
416 drc_mem_usage += sizeof(*rp);
417 }
418
419 /* go ahead and prune the cache */
420 prune_bucket(b);
421
422 found = nfsd_cache_search(b, rqstp, csum);
423 if (found) {
424 if (likely(rp))
425 nfsd_reply_cache_free_locked(rp);
426 rp = found;
427 goto found_entry;
428 }
429
430 if (!rp) {
431 dprintk("nfsd: unable to allocate DRC entry!\n");
432 goto out;
433 }
434
435 nfsdstats.rcmisses++;
436 rqstp->rq_cacherep = rp;
437 rp->c_state = RC_INPROG;
438 rp->c_xid = xid;
439 rp->c_proc = proc;
440 rpc_copy_addr((struct sockaddr *)&rp->c_addr, svc_addr(rqstp));
441 rpc_set_port((struct sockaddr *)&rp->c_addr, rpc_get_port(svc_addr(rqstp)));
442 rp->c_prot = proto;
443 rp->c_vers = vers;
444 rp->c_len = rqstp->rq_arg.len;
445 rp->c_csum = csum;
446
447 lru_put_end(b, rp);
448
449 /* release any buffer */
450 if (rp->c_type == RC_REPLBUFF) {
451 drc_mem_usage -= rp->c_replvec.iov_len;
452 kfree(rp->c_replvec.iov_base);
453 rp->c_replvec.iov_base = NULL;
454 }
455 rp->c_type = RC_NOCACHE;
456 out:
457 spin_unlock(&b->cache_lock);
458 return rtn;
459
460found_entry:
461 nfsdstats.rchits++;
462 /* We found a matching entry which is either in progress or done. */
463 age = jiffies - rp->c_timestamp;
464 lru_put_end(b, rp);
465
466 rtn = RC_DROPIT;
467 /* Request being processed or excessive rexmits */
468 if (rp->c_state == RC_INPROG || age < RC_DELAY)
469 goto out;
470
471 /* From the hall of fame of impractical attacks:
472 * Is this a user who tries to snoop on the cache? */
473 rtn = RC_DOIT;
474 if (!test_bit(RQ_SECURE, &rqstp->rq_flags) && rp->c_secure)
475 goto out;
476
477 /* Compose RPC reply header */
478 switch (rp->c_type) {
479 case RC_NOCACHE:
480 break;
481 case RC_REPLSTAT:
482 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
483 rtn = RC_REPLY;
484 break;
485 case RC_REPLBUFF:
486 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
487 goto out; /* should not happen */
488 rtn = RC_REPLY;
489 break;
490 default:
491 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
492 nfsd_reply_cache_free_locked(rp);
493 }
494
495 goto out;
496}
497
498/*
499 * Update a cache entry. This is called from nfsd_dispatch when
500 * the procedure has been executed and the complete reply is in
501 * rqstp->rq_res.
502 *
503 * We're copying around data here rather than swapping buffers because
504 * the toplevel loop requires max-sized buffers, which would be a waste
505 * of memory for a cache with a max reply size of 100 bytes (diropokres).
506 *
507 * If we should start to use different types of cache entries tailored
508 * specifically for attrstat and fh's, we may save even more space.
509 *
510 * Also note that a cachetype of RC_NOCACHE can legally be passed when
511 * nfsd failed to encode a reply that otherwise would have been cached.
512 * In this case, nfsd_cache_update is called with statp == NULL.
513 */
514void
515nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
516{
517 struct svc_cacherep *rp = rqstp->rq_cacherep;
518 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
519 u32 hash;
520 struct nfsd_drc_bucket *b;
521 int len;
522 size_t bufsize = 0;
523
524 if (!rp)
525 return;
526
527 hash = nfsd_cache_hash(rp->c_xid);
528 b = &drc_hashtbl[hash];
529
530 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
531 len >>= 2;
532
533 /* Don't cache excessive amounts of data and XDR failures */
534 if (!statp || len > (256 >> 2)) {
535 nfsd_reply_cache_free(b, rp);
536 return;
537 }
538
539 switch (cachetype) {
540 case RC_REPLSTAT:
541 if (len != 1)
542 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
543 rp->c_replstat = *statp;
544 break;
545 case RC_REPLBUFF:
546 cachv = &rp->c_replvec;
547 bufsize = len << 2;
548 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
549 if (!cachv->iov_base) {
550 nfsd_reply_cache_free(b, rp);
551 return;
552 }
553 cachv->iov_len = bufsize;
554 memcpy(cachv->iov_base, statp, bufsize);
555 break;
556 case RC_NOCACHE:
557 nfsd_reply_cache_free(b, rp);
558 return;
559 }
560 spin_lock(&b->cache_lock);
561 drc_mem_usage += bufsize;
562 lru_put_end(b, rp);
563 rp->c_secure = test_bit(RQ_SECURE, &rqstp->rq_flags);
564 rp->c_type = cachetype;
565 rp->c_state = RC_DONE;
566 spin_unlock(&b->cache_lock);
567 return;
568}
569
570/*
571 * Copy cached reply to current reply buffer. Should always fit.
572 * FIXME as reply is in a page, we should just attach the page, and
573 * keep a refcount....
574 */
575static int
576nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
577{
578 struct kvec *vec = &rqstp->rq_res.head[0];
579
580 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
581 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
582 data->iov_len);
583 return 0;
584 }
585 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
586 vec->iov_len += data->iov_len;
587 return 1;
588}
589
590/*
591 * Note that fields may be added, removed or reordered in the future. Programs
592 * scraping this file for info should test the labels to ensure they're
593 * getting the correct field.
594 */
595static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
596{
597 seq_printf(m, "max entries: %u\n", max_drc_entries);
598 seq_printf(m, "num entries: %u\n",
599 atomic_read(&num_drc_entries));
600 seq_printf(m, "hash buckets: %u\n", 1 << maskbits);
601 seq_printf(m, "mem usage: %u\n", drc_mem_usage);
602 seq_printf(m, "cache hits: %u\n", nfsdstats.rchits);
603 seq_printf(m, "cache misses: %u\n", nfsdstats.rcmisses);
604 seq_printf(m, "not cached: %u\n", nfsdstats.rcnocache);
605 seq_printf(m, "payload misses: %u\n", payload_misses);
606 seq_printf(m, "longest chain len: %u\n", longest_chain);
607 seq_printf(m, "cachesize at longest: %u\n", longest_chain_cachesize);
608 return 0;
609}
610
611int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
612{
613 return single_open(file, nfsd_reply_cache_stats_show, NULL);
614}