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
30 struct nfsd_drc_bucket {
31 struct list_head lru_head;
32 spinlock_t cache_lock;
33 };
34
35 static struct nfsd_drc_bucket *drc_hashtbl;
36 static struct kmem_cache *drc_slab;
37
38 /* max number of entries allowed in the cache */
39 static unsigned int max_drc_entries;
40
41 /* number of significant bits in the hash value */
42 static unsigned int maskbits;
43 static 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 */
51 static atomic_t num_drc_entries;
52
53 /* cache misses due only to checksum comparison failures */
54 static unsigned int payload_misses;
55
56 /* amount of memory (in bytes) currently consumed by the DRC */
57 static unsigned int drc_mem_usage;
58
59 /* longest hash chain seen */
60 static unsigned int longest_chain;
61
62 /* size of cache when we saw the longest hash chain */
63 static unsigned int longest_chain_cachesize;
64
65 static int nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
66 static void cache_cleaner_func(struct work_struct *unused);
67 static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
68 struct shrink_control *sc);
69 static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
70 struct shrink_control *sc);
71
72 static 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 * locking for the reply cache:
80 * A cache entry is "single use" if c_state == RC_INPROG
81 * Otherwise, it when accessing _prev or _next, the lock must be held.
82 */
83 static DECLARE_DELAYED_WORK(cache_cleaner, cache_cleaner_func);
84
85 /*
86 * Put a cap on the size of the DRC based on the amount of available
87 * low memory in the machine.
88 *
89 * 64MB: 8192
90 * 128MB: 11585
91 * 256MB: 16384
92 * 512MB: 23170
93 * 1GB: 32768
94 * 2GB: 46340
95 * 4GB: 65536
96 * 8GB: 92681
97 * 16GB: 131072
98 *
99 * ...with a hard cap of 256k entries. In the worst case, each entry will be
100 * ~1k, so the above numbers should give a rough max of the amount of memory
101 * used in k.
102 */
103 static unsigned int
nfsd_cache_size_limit(void)104 nfsd_cache_size_limit(void)
105 {
106 unsigned int limit;
107 unsigned long low_pages = totalram_pages - totalhigh_pages;
108
109 limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
110 return min_t(unsigned int, limit, 256*1024);
111 }
112
113 /*
114 * Compute the number of hash buckets we need. Divide the max cachesize by
115 * the "target" max bucket size, and round up to next power of two.
116 */
117 static unsigned int
nfsd_hashsize(unsigned int limit)118 nfsd_hashsize(unsigned int limit)
119 {
120 return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
121 }
122
123 static u32
nfsd_cache_hash(__be32 xid)124 nfsd_cache_hash(__be32 xid)
125 {
126 return hash_32(be32_to_cpu(xid), maskbits);
127 }
128
129 static struct svc_cacherep *
nfsd_reply_cache_alloc(void)130 nfsd_reply_cache_alloc(void)
131 {
132 struct svc_cacherep *rp;
133
134 rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
135 if (rp) {
136 rp->c_state = RC_UNUSED;
137 rp->c_type = RC_NOCACHE;
138 INIT_LIST_HEAD(&rp->c_lru);
139 }
140 return rp;
141 }
142
143 static void
nfsd_reply_cache_free_locked(struct svc_cacherep * rp)144 nfsd_reply_cache_free_locked(struct svc_cacherep *rp)
145 {
146 if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
147 drc_mem_usage -= rp->c_replvec.iov_len;
148 kfree(rp->c_replvec.iov_base);
149 }
150 list_del(&rp->c_lru);
151 atomic_dec(&num_drc_entries);
152 drc_mem_usage -= sizeof(*rp);
153 kmem_cache_free(drc_slab, rp);
154 }
155
156 static void
nfsd_reply_cache_free(struct nfsd_drc_bucket * b,struct svc_cacherep * rp)157 nfsd_reply_cache_free(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
158 {
159 spin_lock(&b->cache_lock);
160 nfsd_reply_cache_free_locked(rp);
161 spin_unlock(&b->cache_lock);
162 }
163
nfsd_reply_cache_init(void)164 int nfsd_reply_cache_init(void)
165 {
166 unsigned int hashsize;
167 unsigned int i;
168
169 max_drc_entries = nfsd_cache_size_limit();
170 atomic_set(&num_drc_entries, 0);
171 hashsize = nfsd_hashsize(max_drc_entries);
172 maskbits = ilog2(hashsize);
173
174 register_shrinker(&nfsd_reply_cache_shrinker);
175 drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
176 0, 0, NULL);
177 if (!drc_slab)
178 goto out_nomem;
179
180 drc_hashtbl = kcalloc(hashsize, sizeof(*drc_hashtbl), GFP_KERNEL);
181 if (!drc_hashtbl)
182 goto out_nomem;
183 for (i = 0; i < hashsize; i++) {
184 INIT_LIST_HEAD(&drc_hashtbl[i].lru_head);
185 spin_lock_init(&drc_hashtbl[i].cache_lock);
186 }
187 drc_hashsize = hashsize;
188
189 return 0;
190 out_nomem:
191 printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
192 nfsd_reply_cache_shutdown();
193 return -ENOMEM;
194 }
195
nfsd_reply_cache_shutdown(void)196 void nfsd_reply_cache_shutdown(void)
197 {
198 struct svc_cacherep *rp;
199 unsigned int i;
200
201 unregister_shrinker(&nfsd_reply_cache_shrinker);
202 cancel_delayed_work_sync(&cache_cleaner);
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 kfree (drc_hashtbl);
213 drc_hashtbl = NULL;
214 drc_hashsize = 0;
215
216 if (drc_slab) {
217 kmem_cache_destroy(drc_slab);
218 drc_slab = NULL;
219 }
220 }
221
222 /*
223 * Move cache entry to end of LRU list, and queue the cleaner to run if it's
224 * not already scheduled.
225 */
226 static void
lru_put_end(struct nfsd_drc_bucket * b,struct svc_cacherep * rp)227 lru_put_end(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
228 {
229 rp->c_timestamp = jiffies;
230 list_move_tail(&rp->c_lru, &b->lru_head);
231 schedule_delayed_work(&cache_cleaner, RC_EXPIRE);
232 }
233
234 static long
prune_bucket(struct nfsd_drc_bucket * b)235 prune_bucket(struct nfsd_drc_bucket *b)
236 {
237 struct svc_cacherep *rp, *tmp;
238 long freed = 0;
239
240 list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
241 /*
242 * Don't free entries attached to calls that are still
243 * in-progress, but do keep scanning the list.
244 */
245 if (rp->c_state == RC_INPROG)
246 continue;
247 if (atomic_read(&num_drc_entries) <= max_drc_entries &&
248 time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
249 break;
250 nfsd_reply_cache_free_locked(rp);
251 freed++;
252 }
253 return freed;
254 }
255
256 /*
257 * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
258 * Also prune the oldest ones when the total exceeds the max number of entries.
259 */
260 static long
prune_cache_entries(void)261 prune_cache_entries(void)
262 {
263 unsigned int i;
264 long freed = 0;
265 bool cancel = true;
266
267 for (i = 0; i < drc_hashsize; i++) {
268 struct nfsd_drc_bucket *b = &drc_hashtbl[i];
269
270 if (list_empty(&b->lru_head))
271 continue;
272 spin_lock(&b->cache_lock);
273 freed += prune_bucket(b);
274 if (!list_empty(&b->lru_head))
275 cancel = false;
276 spin_unlock(&b->cache_lock);
277 }
278
279 /*
280 * Conditionally rearm the job to run in RC_EXPIRE since we just
281 * ran the pruner.
282 */
283 if (!cancel)
284 mod_delayed_work(system_wq, &cache_cleaner, RC_EXPIRE);
285 return freed;
286 }
287
288 static void
cache_cleaner_func(struct work_struct * unused)289 cache_cleaner_func(struct work_struct *unused)
290 {
291 prune_cache_entries();
292 }
293
294 static unsigned long
nfsd_reply_cache_count(struct shrinker * shrink,struct shrink_control * sc)295 nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
296 {
297 return atomic_read(&num_drc_entries);
298 }
299
300 static unsigned long
nfsd_reply_cache_scan(struct shrinker * shrink,struct shrink_control * sc)301 nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
302 {
303 return prune_cache_entries();
304 }
305 /*
306 * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
307 */
308 static __wsum
nfsd_cache_csum(struct svc_rqst * rqstp)309 nfsd_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
338 static bool
nfsd_cache_match(struct svc_rqst * rqstp,__wsum csum,struct svc_cacherep * rp)339 nfsd_cache_match(struct svc_rqst *rqstp, __wsum csum, struct svc_cacherep *rp)
340 {
341 /* Check RPC XID first */
342 if (rqstp->rq_xid != rp->c_xid)
343 return false;
344 /* compare checksum of NFS data */
345 if (csum != rp->c_csum) {
346 ++payload_misses;
347 return false;
348 }
349
350 /* Other discriminators */
351 if (rqstp->rq_proc != rp->c_proc ||
352 rqstp->rq_prot != rp->c_prot ||
353 rqstp->rq_vers != rp->c_vers ||
354 rqstp->rq_arg.len != rp->c_len ||
355 !rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) ||
356 rpc_get_port(svc_addr(rqstp)) != rpc_get_port((struct sockaddr *)&rp->c_addr))
357 return false;
358
359 return true;
360 }
361
362 /*
363 * Search the request hash for an entry that matches the given rqstp.
364 * Must be called with cache_lock held. Returns the found entry or
365 * NULL on failure.
366 */
367 static struct svc_cacherep *
nfsd_cache_search(struct nfsd_drc_bucket * b,struct svc_rqst * rqstp,__wsum csum)368 nfsd_cache_search(struct nfsd_drc_bucket *b, struct svc_rqst *rqstp,
369 __wsum csum)
370 {
371 struct svc_cacherep *rp, *ret = NULL;
372 struct list_head *rh = &b->lru_head;
373 unsigned int entries = 0;
374
375 list_for_each_entry(rp, rh, c_lru) {
376 ++entries;
377 if (nfsd_cache_match(rqstp, csum, rp)) {
378 ret = rp;
379 break;
380 }
381 }
382
383 /* tally hash chain length stats */
384 if (entries > longest_chain) {
385 longest_chain = entries;
386 longest_chain_cachesize = atomic_read(&num_drc_entries);
387 } else if (entries == longest_chain) {
388 /* prefer to keep the smallest cachesize possible here */
389 longest_chain_cachesize = min_t(unsigned int,
390 longest_chain_cachesize,
391 atomic_read(&num_drc_entries));
392 }
393
394 return ret;
395 }
396
397 /*
398 * Try to find an entry matching the current call in the cache. When none
399 * is found, we try to grab the oldest expired entry off the LRU list. If
400 * a suitable one isn't there, then drop the cache_lock and allocate a
401 * new one, then search again in case one got inserted while this thread
402 * didn't hold the lock.
403 */
404 int
nfsd_cache_lookup(struct svc_rqst * rqstp)405 nfsd_cache_lookup(struct svc_rqst *rqstp)
406 {
407 struct svc_cacherep *rp, *found;
408 __be32 xid = rqstp->rq_xid;
409 u32 proto = rqstp->rq_prot,
410 vers = rqstp->rq_vers,
411 proc = rqstp->rq_proc;
412 __wsum csum;
413 u32 hash = nfsd_cache_hash(xid);
414 struct nfsd_drc_bucket *b = &drc_hashtbl[hash];
415 unsigned long age;
416 int type = rqstp->rq_cachetype;
417 int rtn = RC_DOIT;
418
419 rqstp->rq_cacherep = NULL;
420 if (type == RC_NOCACHE) {
421 nfsdstats.rcnocache++;
422 return rtn;
423 }
424
425 csum = nfsd_cache_csum(rqstp);
426
427 /*
428 * Since the common case is a cache miss followed by an insert,
429 * preallocate an entry.
430 */
431 rp = nfsd_reply_cache_alloc();
432 spin_lock(&b->cache_lock);
433 if (likely(rp)) {
434 atomic_inc(&num_drc_entries);
435 drc_mem_usage += sizeof(*rp);
436 }
437
438 /* go ahead and prune the cache */
439 prune_bucket(b);
440
441 found = nfsd_cache_search(b, rqstp, csum);
442 if (found) {
443 if (likely(rp))
444 nfsd_reply_cache_free_locked(rp);
445 rp = found;
446 goto found_entry;
447 }
448
449 if (!rp) {
450 dprintk("nfsd: unable to allocate DRC entry!\n");
451 goto out;
452 }
453
454 nfsdstats.rcmisses++;
455 rqstp->rq_cacherep = rp;
456 rp->c_state = RC_INPROG;
457 rp->c_xid = xid;
458 rp->c_proc = proc;
459 rpc_copy_addr((struct sockaddr *)&rp->c_addr, svc_addr(rqstp));
460 rpc_set_port((struct sockaddr *)&rp->c_addr, rpc_get_port(svc_addr(rqstp)));
461 rp->c_prot = proto;
462 rp->c_vers = vers;
463 rp->c_len = rqstp->rq_arg.len;
464 rp->c_csum = csum;
465
466 lru_put_end(b, rp);
467
468 /* release any buffer */
469 if (rp->c_type == RC_REPLBUFF) {
470 drc_mem_usage -= rp->c_replvec.iov_len;
471 kfree(rp->c_replvec.iov_base);
472 rp->c_replvec.iov_base = NULL;
473 }
474 rp->c_type = RC_NOCACHE;
475 out:
476 spin_unlock(&b->cache_lock);
477 return rtn;
478
479 found_entry:
480 nfsdstats.rchits++;
481 /* We found a matching entry which is either in progress or done. */
482 age = jiffies - rp->c_timestamp;
483 lru_put_end(b, rp);
484
485 rtn = RC_DROPIT;
486 /* Request being processed or excessive rexmits */
487 if (rp->c_state == RC_INPROG || age < RC_DELAY)
488 goto out;
489
490 /* From the hall of fame of impractical attacks:
491 * Is this a user who tries to snoop on the cache? */
492 rtn = RC_DOIT;
493 if (!rqstp->rq_secure && rp->c_secure)
494 goto out;
495
496 /* Compose RPC reply header */
497 switch (rp->c_type) {
498 case RC_NOCACHE:
499 break;
500 case RC_REPLSTAT:
501 svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
502 rtn = RC_REPLY;
503 break;
504 case RC_REPLBUFF:
505 if (!nfsd_cache_append(rqstp, &rp->c_replvec))
506 goto out; /* should not happen */
507 rtn = RC_REPLY;
508 break;
509 default:
510 printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
511 nfsd_reply_cache_free_locked(rp);
512 }
513
514 goto out;
515 }
516
517 /*
518 * Update a cache entry. This is called from nfsd_dispatch when
519 * the procedure has been executed and the complete reply is in
520 * rqstp->rq_res.
521 *
522 * We're copying around data here rather than swapping buffers because
523 * the toplevel loop requires max-sized buffers, which would be a waste
524 * of memory for a cache with a max reply size of 100 bytes (diropokres).
525 *
526 * If we should start to use different types of cache entries tailored
527 * specifically for attrstat and fh's, we may save even more space.
528 *
529 * Also note that a cachetype of RC_NOCACHE can legally be passed when
530 * nfsd failed to encode a reply that otherwise would have been cached.
531 * In this case, nfsd_cache_update is called with statp == NULL.
532 */
533 void
nfsd_cache_update(struct svc_rqst * rqstp,int cachetype,__be32 * statp)534 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
535 {
536 struct svc_cacherep *rp = rqstp->rq_cacherep;
537 struct kvec *resv = &rqstp->rq_res.head[0], *cachv;
538 u32 hash;
539 struct nfsd_drc_bucket *b;
540 int len;
541 size_t bufsize = 0;
542
543 if (!rp)
544 return;
545
546 hash = nfsd_cache_hash(rp->c_xid);
547 b = &drc_hashtbl[hash];
548
549 len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
550 len >>= 2;
551
552 /* Don't cache excessive amounts of data and XDR failures */
553 if (!statp || len > (256 >> 2)) {
554 nfsd_reply_cache_free(b, rp);
555 return;
556 }
557
558 switch (cachetype) {
559 case RC_REPLSTAT:
560 if (len != 1)
561 printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
562 rp->c_replstat = *statp;
563 break;
564 case RC_REPLBUFF:
565 cachv = &rp->c_replvec;
566 bufsize = len << 2;
567 cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
568 if (!cachv->iov_base) {
569 nfsd_reply_cache_free(b, rp);
570 return;
571 }
572 cachv->iov_len = bufsize;
573 memcpy(cachv->iov_base, statp, bufsize);
574 break;
575 case RC_NOCACHE:
576 nfsd_reply_cache_free(b, rp);
577 return;
578 }
579 spin_lock(&b->cache_lock);
580 drc_mem_usage += bufsize;
581 lru_put_end(b, rp);
582 rp->c_secure = rqstp->rq_secure;
583 rp->c_type = cachetype;
584 rp->c_state = RC_DONE;
585 spin_unlock(&b->cache_lock);
586 return;
587 }
588
589 /*
590 * Copy cached reply to current reply buffer. Should always fit.
591 * FIXME as reply is in a page, we should just attach the page, and
592 * keep a refcount....
593 */
594 static int
nfsd_cache_append(struct svc_rqst * rqstp,struct kvec * data)595 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
596 {
597 struct kvec *vec = &rqstp->rq_res.head[0];
598
599 if (vec->iov_len + data->iov_len > PAGE_SIZE) {
600 printk(KERN_WARNING "nfsd: cached reply too large (%Zd).\n",
601 data->iov_len);
602 return 0;
603 }
604 memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
605 vec->iov_len += data->iov_len;
606 return 1;
607 }
608
609 /*
610 * Note that fields may be added, removed or reordered in the future. Programs
611 * scraping this file for info should test the labels to ensure they're
612 * getting the correct field.
613 */
nfsd_reply_cache_stats_show(struct seq_file * m,void * v)614 static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
615 {
616 seq_printf(m, "max entries: %u\n", max_drc_entries);
617 seq_printf(m, "num entries: %u\n",
618 atomic_read(&num_drc_entries));
619 seq_printf(m, "hash buckets: %u\n", 1 << maskbits);
620 seq_printf(m, "mem usage: %u\n", drc_mem_usage);
621 seq_printf(m, "cache hits: %u\n", nfsdstats.rchits);
622 seq_printf(m, "cache misses: %u\n", nfsdstats.rcmisses);
623 seq_printf(m, "not cached: %u\n", nfsdstats.rcnocache);
624 seq_printf(m, "payload misses: %u\n", payload_misses);
625 seq_printf(m, "longest chain len: %u\n", longest_chain);
626 seq_printf(m, "cachesize at longest: %u\n", longest_chain_cachesize);
627 return 0;
628 }
629
nfsd_reply_cache_stats_open(struct inode * inode,struct file * file)630 int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
631 {
632 return single_open(file, nfsd_reply_cache_stats_show, NULL);
633 }
634