xref: /openbmc/linux/fs/nfsd/nfscache.c (revision 2359ccdd)
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/slab.h>
13 #include <linux/vmalloc.h>
14 #include <linux/sunrpc/addr.h>
15 #include <linux/highmem.h>
16 #include <linux/log2.h>
17 #include <linux/hash.h>
18 #include <net/checksum.h>
19 
20 #include "nfsd.h"
21 #include "cache.h"
22 
23 #define NFSDDBG_FACILITY	NFSDDBG_REPCACHE
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 
32 struct nfsd_drc_bucket {
33 	struct list_head lru_head;
34 	spinlock_t cache_lock;
35 };
36 
37 static struct nfsd_drc_bucket	*drc_hashtbl;
38 static struct kmem_cache	*drc_slab;
39 
40 /* max number of entries allowed in the cache */
41 static unsigned int		max_drc_entries;
42 
43 /* number of significant bits in the hash value */
44 static unsigned int		maskbits;
45 static unsigned int		drc_hashsize;
46 
47 /*
48  * Stats and other tracking of on the duplicate reply cache. All of these and
49  * the "rc" fields in nfsdstats are protected by the cache_lock
50  */
51 
52 /* total number of entries */
53 static atomic_t			num_drc_entries;
54 
55 /* cache misses due only to checksum comparison failures */
56 static unsigned int		payload_misses;
57 
58 /* amount of memory (in bytes) currently consumed by the DRC */
59 static unsigned int		drc_mem_usage;
60 
61 /* longest hash chain seen */
62 static unsigned int		longest_chain;
63 
64 /* size of cache when we saw the longest hash chain */
65 static unsigned int		longest_chain_cachesize;
66 
67 static int	nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
68 static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
69 					    struct shrink_control *sc);
70 static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
71 					   struct shrink_control *sc);
72 
73 static struct shrinker nfsd_reply_cache_shrinker = {
74 	.scan_objects = nfsd_reply_cache_scan,
75 	.count_objects = nfsd_reply_cache_count,
76 	.seeks	= 1,
77 };
78 
79 /*
80  * Put a cap on the size of the DRC based on the amount of available
81  * low memory in the machine.
82  *
83  *  64MB:    8192
84  * 128MB:   11585
85  * 256MB:   16384
86  * 512MB:   23170
87  *   1GB:   32768
88  *   2GB:   46340
89  *   4GB:   65536
90  *   8GB:   92681
91  *  16GB:  131072
92  *
93  * ...with a hard cap of 256k entries. In the worst case, each entry will be
94  * ~1k, so the above numbers should give a rough max of the amount of memory
95  * used in k.
96  */
97 static unsigned int
98 nfsd_cache_size_limit(void)
99 {
100 	unsigned int limit;
101 	unsigned long low_pages = totalram_pages - totalhigh_pages;
102 
103 	limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
104 	return min_t(unsigned int, limit, 256*1024);
105 }
106 
107 /*
108  * Compute the number of hash buckets we need. Divide the max cachesize by
109  * the "target" max bucket size, and round up to next power of two.
110  */
111 static unsigned int
112 nfsd_hashsize(unsigned int limit)
113 {
114 	return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
115 }
116 
117 static u32
118 nfsd_cache_hash(__be32 xid)
119 {
120 	return hash_32(be32_to_cpu(xid), maskbits);
121 }
122 
123 static struct svc_cacherep *
124 nfsd_reply_cache_alloc(void)
125 {
126 	struct svc_cacherep	*rp;
127 
128 	rp = kmem_cache_alloc(drc_slab, GFP_KERNEL);
129 	if (rp) {
130 		rp->c_state = RC_UNUSED;
131 		rp->c_type = RC_NOCACHE;
132 		INIT_LIST_HEAD(&rp->c_lru);
133 	}
134 	return rp;
135 }
136 
137 static void
138 nfsd_reply_cache_free_locked(struct svc_cacherep *rp)
139 {
140 	if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
141 		drc_mem_usage -= rp->c_replvec.iov_len;
142 		kfree(rp->c_replvec.iov_base);
143 	}
144 	list_del(&rp->c_lru);
145 	atomic_dec(&num_drc_entries);
146 	drc_mem_usage -= sizeof(*rp);
147 	kmem_cache_free(drc_slab, rp);
148 }
149 
150 static void
151 nfsd_reply_cache_free(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
152 {
153 	spin_lock(&b->cache_lock);
154 	nfsd_reply_cache_free_locked(rp);
155 	spin_unlock(&b->cache_lock);
156 }
157 
158 int nfsd_reply_cache_init(void)
159 {
160 	unsigned int hashsize;
161 	unsigned int i;
162 	int status = 0;
163 
164 	max_drc_entries = nfsd_cache_size_limit();
165 	atomic_set(&num_drc_entries, 0);
166 	hashsize = nfsd_hashsize(max_drc_entries);
167 	maskbits = ilog2(hashsize);
168 
169 	status = register_shrinker(&nfsd_reply_cache_shrinker);
170 	if (status)
171 		return status;
172 
173 	drc_slab = kmem_cache_create("nfsd_drc", sizeof(struct svc_cacherep),
174 					0, 0, NULL);
175 	if (!drc_slab)
176 		goto out_nomem;
177 
178 	drc_hashtbl = kcalloc(hashsize, sizeof(*drc_hashtbl), GFP_KERNEL);
179 	if (!drc_hashtbl) {
180 		drc_hashtbl = vzalloc(hashsize * sizeof(*drc_hashtbl));
181 		if (!drc_hashtbl)
182 			goto out_nomem;
183 	}
184 
185 	for (i = 0; i < hashsize; i++) {
186 		INIT_LIST_HEAD(&drc_hashtbl[i].lru_head);
187 		spin_lock_init(&drc_hashtbl[i].cache_lock);
188 	}
189 	drc_hashsize = hashsize;
190 
191 	return 0;
192 out_nomem:
193 	printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
194 	nfsd_reply_cache_shutdown();
195 	return -ENOMEM;
196 }
197 
198 void nfsd_reply_cache_shutdown(void)
199 {
200 	struct svc_cacherep	*rp;
201 	unsigned int i;
202 
203 	unregister_shrinker(&nfsd_reply_cache_shrinker);
204 
205 	for (i = 0; i < drc_hashsize; i++) {
206 		struct list_head *head = &drc_hashtbl[i].lru_head;
207 		while (!list_empty(head)) {
208 			rp = list_first_entry(head, struct svc_cacherep, c_lru);
209 			nfsd_reply_cache_free_locked(rp);
210 		}
211 	}
212 
213 	kvfree(drc_hashtbl);
214 	drc_hashtbl = NULL;
215 	drc_hashsize = 0;
216 
217 	kmem_cache_destroy(drc_slab);
218 	drc_slab = NULL;
219 }
220 
221 /*
222  * Move cache entry to end of LRU list, and queue the cleaner to run if it's
223  * not already scheduled.
224  */
225 static void
226 lru_put_end(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
227 {
228 	rp->c_timestamp = jiffies;
229 	list_move_tail(&rp->c_lru, &b->lru_head);
230 }
231 
232 static long
233 prune_bucket(struct nfsd_drc_bucket *b)
234 {
235 	struct svc_cacherep *rp, *tmp;
236 	long freed = 0;
237 
238 	list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
239 		/*
240 		 * Don't free entries attached to calls that are still
241 		 * in-progress, but do keep scanning the list.
242 		 */
243 		if (rp->c_state == RC_INPROG)
244 			continue;
245 		if (atomic_read(&num_drc_entries) <= max_drc_entries &&
246 		    time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
247 			break;
248 		nfsd_reply_cache_free_locked(rp);
249 		freed++;
250 	}
251 	return freed;
252 }
253 
254 /*
255  * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
256  * Also prune the oldest ones when the total exceeds the max number of entries.
257  */
258 static long
259 prune_cache_entries(void)
260 {
261 	unsigned int i;
262 	long freed = 0;
263 
264 	for (i = 0; i < drc_hashsize; i++) {
265 		struct nfsd_drc_bucket *b = &drc_hashtbl[i];
266 
267 		if (list_empty(&b->lru_head))
268 			continue;
269 		spin_lock(&b->cache_lock);
270 		freed += prune_bucket(b);
271 		spin_unlock(&b->cache_lock);
272 	}
273 	return freed;
274 }
275 
276 static unsigned long
277 nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
278 {
279 	return atomic_read(&num_drc_entries);
280 }
281 
282 static unsigned long
283 nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
284 {
285 	return prune_cache_entries();
286 }
287 /*
288  * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
289  */
290 static __wsum
291 nfsd_cache_csum(struct svc_rqst *rqstp)
292 {
293 	int idx;
294 	unsigned int base;
295 	__wsum csum;
296 	struct xdr_buf *buf = &rqstp->rq_arg;
297 	const unsigned char *p = buf->head[0].iov_base;
298 	size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
299 				RC_CSUMLEN);
300 	size_t len = min(buf->head[0].iov_len, csum_len);
301 
302 	/* rq_arg.head first */
303 	csum = csum_partial(p, len, 0);
304 	csum_len -= len;
305 
306 	/* Continue into page array */
307 	idx = buf->page_base / PAGE_SIZE;
308 	base = buf->page_base & ~PAGE_MASK;
309 	while (csum_len) {
310 		p = page_address(buf->pages[idx]) + base;
311 		len = min_t(size_t, PAGE_SIZE - base, csum_len);
312 		csum = csum_partial(p, len, csum);
313 		csum_len -= len;
314 		base = 0;
315 		++idx;
316 	}
317 	return csum;
318 }
319 
320 static bool
321 nfsd_cache_match(struct svc_rqst *rqstp, __wsum csum, struct svc_cacherep *rp)
322 {
323 	/* Check RPC XID first */
324 	if (rqstp->rq_xid != rp->c_xid)
325 		return false;
326 	/* compare checksum of NFS data */
327 	if (csum != rp->c_csum) {
328 		++payload_misses;
329 		return false;
330 	}
331 
332 	/* Other discriminators */
333 	if (rqstp->rq_proc != rp->c_proc ||
334 	    rqstp->rq_prot != rp->c_prot ||
335 	    rqstp->rq_vers != rp->c_vers ||
336 	    rqstp->rq_arg.len != rp->c_len ||
337 	    !rpc_cmp_addr(svc_addr(rqstp), (struct sockaddr *)&rp->c_addr) ||
338 	    rpc_get_port(svc_addr(rqstp)) != rpc_get_port((struct sockaddr *)&rp->c_addr))
339 		return false;
340 
341 	return true;
342 }
343 
344 /*
345  * Search the request hash for an entry that matches the given rqstp.
346  * Must be called with cache_lock held. Returns the found entry or
347  * NULL on failure.
348  */
349 static struct svc_cacherep *
350 nfsd_cache_search(struct nfsd_drc_bucket *b, struct svc_rqst *rqstp,
351 		__wsum csum)
352 {
353 	struct svc_cacherep	*rp, *ret = NULL;
354 	struct list_head 	*rh = &b->lru_head;
355 	unsigned int		entries = 0;
356 
357 	list_for_each_entry(rp, rh, c_lru) {
358 		++entries;
359 		if (nfsd_cache_match(rqstp, csum, rp)) {
360 			ret = rp;
361 			break;
362 		}
363 	}
364 
365 	/* tally hash chain length stats */
366 	if (entries > longest_chain) {
367 		longest_chain = entries;
368 		longest_chain_cachesize = atomic_read(&num_drc_entries);
369 	} else if (entries == longest_chain) {
370 		/* prefer to keep the smallest cachesize possible here */
371 		longest_chain_cachesize = min_t(unsigned int,
372 				longest_chain_cachesize,
373 				atomic_read(&num_drc_entries));
374 	}
375 
376 	return ret;
377 }
378 
379 /*
380  * Try to find an entry matching the current call in the cache. When none
381  * is found, we try to grab the oldest expired entry off the LRU list. If
382  * a suitable one isn't there, then drop the cache_lock and allocate a
383  * new one, then search again in case one got inserted while this thread
384  * didn't hold the lock.
385  */
386 int
387 nfsd_cache_lookup(struct svc_rqst *rqstp)
388 {
389 	struct svc_cacherep	*rp, *found;
390 	__be32			xid = rqstp->rq_xid;
391 	u32			proto =  rqstp->rq_prot,
392 				vers = rqstp->rq_vers,
393 				proc = rqstp->rq_proc;
394 	__wsum			csum;
395 	u32 hash = nfsd_cache_hash(xid);
396 	struct nfsd_drc_bucket *b = &drc_hashtbl[hash];
397 	unsigned long		age;
398 	int type = rqstp->rq_cachetype;
399 	int rtn = RC_DOIT;
400 
401 	rqstp->rq_cacherep = NULL;
402 	if (type == RC_NOCACHE) {
403 		nfsdstats.rcnocache++;
404 		return rtn;
405 	}
406 
407 	csum = nfsd_cache_csum(rqstp);
408 
409 	/*
410 	 * Since the common case is a cache miss followed by an insert,
411 	 * preallocate an entry.
412 	 */
413 	rp = nfsd_reply_cache_alloc();
414 	spin_lock(&b->cache_lock);
415 	if (likely(rp)) {
416 		atomic_inc(&num_drc_entries);
417 		drc_mem_usage += sizeof(*rp);
418 	}
419 
420 	/* go ahead and prune the cache */
421 	prune_bucket(b);
422 
423 	found = nfsd_cache_search(b, rqstp, csum);
424 	if (found) {
425 		if (likely(rp))
426 			nfsd_reply_cache_free_locked(rp);
427 		rp = found;
428 		goto found_entry;
429 	}
430 
431 	if (!rp) {
432 		dprintk("nfsd: unable to allocate DRC entry!\n");
433 		goto out;
434 	}
435 
436 	nfsdstats.rcmisses++;
437 	rqstp->rq_cacherep = rp;
438 	rp->c_state = RC_INPROG;
439 	rp->c_xid = xid;
440 	rp->c_proc = proc;
441 	rpc_copy_addr((struct sockaddr *)&rp->c_addr, svc_addr(rqstp));
442 	rpc_set_port((struct sockaddr *)&rp->c_addr, rpc_get_port(svc_addr(rqstp)));
443 	rp->c_prot = proto;
444 	rp->c_vers = vers;
445 	rp->c_len = rqstp->rq_arg.len;
446 	rp->c_csum = csum;
447 
448 	lru_put_end(b, rp);
449 
450 	/* release any buffer */
451 	if (rp->c_type == RC_REPLBUFF) {
452 		drc_mem_usage -= rp->c_replvec.iov_len;
453 		kfree(rp->c_replvec.iov_base);
454 		rp->c_replvec.iov_base = NULL;
455 	}
456 	rp->c_type = RC_NOCACHE;
457  out:
458 	spin_unlock(&b->cache_lock);
459 	return rtn;
460 
461 found_entry:
462 	nfsdstats.rchits++;
463 	/* We found a matching entry which is either in progress or done. */
464 	age = jiffies - rp->c_timestamp;
465 	lru_put_end(b, rp);
466 
467 	rtn = RC_DROPIT;
468 	/* Request being processed or excessive rexmits */
469 	if (rp->c_state == RC_INPROG || age < RC_DELAY)
470 		goto out;
471 
472 	/* From the hall of fame of impractical attacks:
473 	 * Is this a user who tries to snoop on the cache? */
474 	rtn = RC_DOIT;
475 	if (!test_bit(RQ_SECURE, &rqstp->rq_flags) && rp->c_secure)
476 		goto out;
477 
478 	/* Compose RPC reply header */
479 	switch (rp->c_type) {
480 	case RC_NOCACHE:
481 		break;
482 	case RC_REPLSTAT:
483 		svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
484 		rtn = RC_REPLY;
485 		break;
486 	case RC_REPLBUFF:
487 		if (!nfsd_cache_append(rqstp, &rp->c_replvec))
488 			goto out;	/* should not happen */
489 		rtn = RC_REPLY;
490 		break;
491 	default:
492 		printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
493 		nfsd_reply_cache_free_locked(rp);
494 	}
495 
496 	goto out;
497 }
498 
499 /*
500  * Update a cache entry. This is called from nfsd_dispatch when
501  * the procedure has been executed and the complete reply is in
502  * rqstp->rq_res.
503  *
504  * We're copying around data here rather than swapping buffers because
505  * the toplevel loop requires max-sized buffers, which would be a waste
506  * of memory for a cache with a max reply size of 100 bytes (diropokres).
507  *
508  * If we should start to use different types of cache entries tailored
509  * specifically for attrstat and fh's, we may save even more space.
510  *
511  * Also note that a cachetype of RC_NOCACHE can legally be passed when
512  * nfsd failed to encode a reply that otherwise would have been cached.
513  * In this case, nfsd_cache_update is called with statp == NULL.
514  */
515 void
516 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
517 {
518 	struct svc_cacherep *rp = rqstp->rq_cacherep;
519 	struct kvec	*resv = &rqstp->rq_res.head[0], *cachv;
520 	u32		hash;
521 	struct nfsd_drc_bucket *b;
522 	int		len;
523 	size_t		bufsize = 0;
524 
525 	if (!rp)
526 		return;
527 
528 	hash = nfsd_cache_hash(rp->c_xid);
529 	b = &drc_hashtbl[hash];
530 
531 	len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
532 	len >>= 2;
533 
534 	/* Don't cache excessive amounts of data and XDR failures */
535 	if (!statp || len > (256 >> 2)) {
536 		nfsd_reply_cache_free(b, rp);
537 		return;
538 	}
539 
540 	switch (cachetype) {
541 	case RC_REPLSTAT:
542 		if (len != 1)
543 			printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
544 		rp->c_replstat = *statp;
545 		break;
546 	case RC_REPLBUFF:
547 		cachv = &rp->c_replvec;
548 		bufsize = len << 2;
549 		cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
550 		if (!cachv->iov_base) {
551 			nfsd_reply_cache_free(b, rp);
552 			return;
553 		}
554 		cachv->iov_len = bufsize;
555 		memcpy(cachv->iov_base, statp, bufsize);
556 		break;
557 	case RC_NOCACHE:
558 		nfsd_reply_cache_free(b, rp);
559 		return;
560 	}
561 	spin_lock(&b->cache_lock);
562 	drc_mem_usage += bufsize;
563 	lru_put_end(b, rp);
564 	rp->c_secure = test_bit(RQ_SECURE, &rqstp->rq_flags);
565 	rp->c_type = cachetype;
566 	rp->c_state = RC_DONE;
567 	spin_unlock(&b->cache_lock);
568 	return;
569 }
570 
571 /*
572  * Copy cached reply to current reply buffer. Should always fit.
573  * FIXME as reply is in a page, we should just attach the page, and
574  * keep a refcount....
575  */
576 static int
577 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
578 {
579 	struct kvec	*vec = &rqstp->rq_res.head[0];
580 
581 	if (vec->iov_len + data->iov_len > PAGE_SIZE) {
582 		printk(KERN_WARNING "nfsd: cached reply too large (%zd).\n",
583 				data->iov_len);
584 		return 0;
585 	}
586 	memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
587 	vec->iov_len += data->iov_len;
588 	return 1;
589 }
590 
591 /*
592  * Note that fields may be added, removed or reordered in the future. Programs
593  * scraping this file for info should test the labels to ensure they're
594  * getting the correct field.
595  */
596 static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
597 {
598 	seq_printf(m, "max entries:           %u\n", max_drc_entries);
599 	seq_printf(m, "num entries:           %u\n",
600 			atomic_read(&num_drc_entries));
601 	seq_printf(m, "hash buckets:          %u\n", 1 << maskbits);
602 	seq_printf(m, "mem usage:             %u\n", drc_mem_usage);
603 	seq_printf(m, "cache hits:            %u\n", nfsdstats.rchits);
604 	seq_printf(m, "cache misses:          %u\n", nfsdstats.rcmisses);
605 	seq_printf(m, "not cached:            %u\n", nfsdstats.rcnocache);
606 	seq_printf(m, "payload misses:        %u\n", payload_misses);
607 	seq_printf(m, "longest chain len:     %u\n", longest_chain);
608 	seq_printf(m, "cachesize at longest:  %u\n", longest_chain_cachesize);
609 	return 0;
610 }
611 
612 int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
613 {
614 	return single_open(file, nfsd_reply_cache_stats_show, NULL);
615 }
616