xref: /openbmc/linux/fs/nfsd/nfscache.c (revision 3805e6a1)
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 unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
67 					    struct shrink_control *sc);
68 static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
69 					   struct shrink_control *sc);
70 
71 static 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  */
95 static unsigned int
96 nfsd_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  */
109 static unsigned int
110 nfsd_hashsize(unsigned int limit)
111 {
112 	return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
113 }
114 
115 static u32
116 nfsd_cache_hash(__be32 xid)
117 {
118 	return hash_32(be32_to_cpu(xid), maskbits);
119 }
120 
121 static struct svc_cacherep *
122 nfsd_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 
135 static void
136 nfsd_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 
148 static void
149 nfsd_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 
156 int 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;
186 out_nomem:
187 	printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
188 	nfsd_reply_cache_shutdown();
189 	return -ENOMEM;
190 }
191 
192 void 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  */
219 static void
220 lru_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 
226 static long
227 prune_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  */
252 static long
253 prune_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 
270 static unsigned long
271 nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
272 {
273 	return atomic_read(&num_drc_entries);
274 }
275 
276 static unsigned long
277 nfsd_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  */
284 static __wsum
285 nfsd_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 
314 static bool
315 nfsd_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  */
343 static struct svc_cacherep *
344 nfsd_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  */
380 int
381 nfsd_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 
455 found_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  */
509 void
510 nfsd_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  */
570 static int
571 nfsd_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  */
590 static 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 
606 int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
607 {
608 	return single_open(file, nfsd_reply_cache_stats_show, NULL);
609 }
610