xref: /openbmc/linux/fs/nfsd/nfscache.c (revision b34e08d5)
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 static struct hlist_head *	cache_hash;
31 static struct list_head 	lru_head;
32 static struct kmem_cache	*drc_slab;
33 
34 /* max number of entries allowed in the cache */
35 static unsigned int		max_drc_entries;
36 
37 /* number of significant bits in the hash value */
38 static 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 */
46 static unsigned int		num_drc_entries;
47 
48 /* cache misses due only to checksum comparison failures */
49 static unsigned int		payload_misses;
50 
51 /* amount of memory (in bytes) currently consumed by the DRC */
52 static unsigned int		drc_mem_usage;
53 
54 /* longest hash chain seen */
55 static unsigned int		longest_chain;
56 
57 /* size of cache when we saw the longest hash chain */
58 static unsigned int		longest_chain_cachesize;
59 
60 static int	nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
61 static void	cache_cleaner_func(struct work_struct *unused);
62 static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
63 					    struct shrink_control *sc);
64 static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
65 					   struct shrink_control *sc);
66 
67 static 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  */
78 static DEFINE_SPINLOCK(cache_lock);
79 static 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  */
99 static unsigned int
100 nfsd_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  */
113 static unsigned int
114 nfsd_hashsize(unsigned int limit)
115 {
116 	return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
117 }
118 
119 static struct svc_cacherep *
120 nfsd_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 
134 static void
135 nfsd_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 
149 static void
150 nfsd_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 
157 int 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;
178 out_nomem:
179 	printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
180 	nfsd_reply_cache_shutdown();
181 	return -ENOMEM;
182 }
183 
184 void 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  */
209 static void
210 lru_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  */
220 static void
221 hash_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 
227 static inline bool
228 nfsd_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  */
238 static long
239 prune_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 
265 static void
266 cache_cleaner_func(struct work_struct *unused)
267 {
268 	spin_lock(&cache_lock);
269 	prune_cache_entries();
270 	spin_unlock(&cache_lock);
271 }
272 
273 static unsigned long
274 nfsd_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 
285 static unsigned long
286 nfsd_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  */
298 static __wsum
299 nfsd_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 
328 static bool
329 nfsd_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  */
353 static struct svc_cacherep *
354 nfsd_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  */
389 int
390 nfsd_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 
463 found_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  */
517 void
518 nfsd_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  */
573 static int
574 nfsd_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  */
593 static 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 
610 int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
611 {
612 	return single_open(file, nfsd_reply_cache_stats_show, NULL);
613 }
614