xref: /openbmc/linux/fs/nfsd/nfscache.c (revision f8e17c17)
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/sunrpc/svc_xprt.h>
13 #include <linux/slab.h>
14 #include <linux/vmalloc.h>
15 #include <linux/sunrpc/addr.h>
16 #include <linux/highmem.h>
17 #include <linux/log2.h>
18 #include <linux/hash.h>
19 #include <net/checksum.h>
20 
21 #include "nfsd.h"
22 #include "cache.h"
23 
24 #define NFSDDBG_FACILITY	NFSDDBG_REPCACHE
25 
26 /*
27  * We use this value to determine the number of hash buckets from the max
28  * cache size, the idea being that when the cache is at its maximum number
29  * of entries, then this should be the average number of entries per bucket.
30  */
31 #define TARGET_BUCKET_SIZE	64
32 
33 struct nfsd_drc_bucket {
34 	struct rb_root rb_head;
35 	struct list_head lru_head;
36 	spinlock_t cache_lock;
37 };
38 
39 static int	nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *vec);
40 static unsigned long nfsd_reply_cache_count(struct shrinker *shrink,
41 					    struct shrink_control *sc);
42 static unsigned long nfsd_reply_cache_scan(struct shrinker *shrink,
43 					   struct shrink_control *sc);
44 
45 /*
46  * Put a cap on the size of the DRC based on the amount of available
47  * low memory in the machine.
48  *
49  *  64MB:    8192
50  * 128MB:   11585
51  * 256MB:   16384
52  * 512MB:   23170
53  *   1GB:   32768
54  *   2GB:   46340
55  *   4GB:   65536
56  *   8GB:   92681
57  *  16GB:  131072
58  *
59  * ...with a hard cap of 256k entries. In the worst case, each entry will be
60  * ~1k, so the above numbers should give a rough max of the amount of memory
61  * used in k.
62  *
63  * XXX: these limits are per-container, so memory used will increase
64  * linearly with number of containers.  Maybe that's OK.
65  */
66 static unsigned int
67 nfsd_cache_size_limit(void)
68 {
69 	unsigned int limit;
70 	unsigned long low_pages = totalram_pages() - totalhigh_pages();
71 
72 	limit = (16 * int_sqrt(low_pages)) << (PAGE_SHIFT-10);
73 	return min_t(unsigned int, limit, 256*1024);
74 }
75 
76 /*
77  * Compute the number of hash buckets we need. Divide the max cachesize by
78  * the "target" max bucket size, and round up to next power of two.
79  */
80 static unsigned int
81 nfsd_hashsize(unsigned int limit)
82 {
83 	return roundup_pow_of_two(limit / TARGET_BUCKET_SIZE);
84 }
85 
86 static u32
87 nfsd_cache_hash(__be32 xid, struct nfsd_net *nn)
88 {
89 	return hash_32(be32_to_cpu(xid), nn->maskbits);
90 }
91 
92 static struct svc_cacherep *
93 nfsd_reply_cache_alloc(struct svc_rqst *rqstp, __wsum csum,
94 			struct nfsd_net *nn)
95 {
96 	struct svc_cacherep	*rp;
97 
98 	rp = kmem_cache_alloc(nn->drc_slab, GFP_KERNEL);
99 	if (rp) {
100 		rp->c_state = RC_UNUSED;
101 		rp->c_type = RC_NOCACHE;
102 		RB_CLEAR_NODE(&rp->c_node);
103 		INIT_LIST_HEAD(&rp->c_lru);
104 
105 		memset(&rp->c_key, 0, sizeof(rp->c_key));
106 		rp->c_key.k_xid = rqstp->rq_xid;
107 		rp->c_key.k_proc = rqstp->rq_proc;
108 		rpc_copy_addr((struct sockaddr *)&rp->c_key.k_addr, svc_addr(rqstp));
109 		rpc_set_port((struct sockaddr *)&rp->c_key.k_addr, rpc_get_port(svc_addr(rqstp)));
110 		rp->c_key.k_prot = rqstp->rq_prot;
111 		rp->c_key.k_vers = rqstp->rq_vers;
112 		rp->c_key.k_len = rqstp->rq_arg.len;
113 		rp->c_key.k_csum = csum;
114 	}
115 	return rp;
116 }
117 
118 static void
119 nfsd_reply_cache_free_locked(struct nfsd_drc_bucket *b, struct svc_cacherep *rp,
120 				struct nfsd_net *nn)
121 {
122 	if (rp->c_type == RC_REPLBUFF && rp->c_replvec.iov_base) {
123 		nn->drc_mem_usage -= rp->c_replvec.iov_len;
124 		kfree(rp->c_replvec.iov_base);
125 	}
126 	if (rp->c_state != RC_UNUSED) {
127 		rb_erase(&rp->c_node, &b->rb_head);
128 		list_del(&rp->c_lru);
129 		atomic_dec(&nn->num_drc_entries);
130 		nn->drc_mem_usage -= sizeof(*rp);
131 	}
132 	kmem_cache_free(nn->drc_slab, rp);
133 }
134 
135 static void
136 nfsd_reply_cache_free(struct nfsd_drc_bucket *b, struct svc_cacherep *rp,
137 			struct nfsd_net *nn)
138 {
139 	spin_lock(&b->cache_lock);
140 	nfsd_reply_cache_free_locked(b, rp, nn);
141 	spin_unlock(&b->cache_lock);
142 }
143 
144 int nfsd_reply_cache_init(struct nfsd_net *nn)
145 {
146 	unsigned int hashsize;
147 	unsigned int i;
148 	int status = 0;
149 
150 	nn->max_drc_entries = nfsd_cache_size_limit();
151 	atomic_set(&nn->num_drc_entries, 0);
152 	hashsize = nfsd_hashsize(nn->max_drc_entries);
153 	nn->maskbits = ilog2(hashsize);
154 
155 	nn->nfsd_reply_cache_shrinker.scan_objects = nfsd_reply_cache_scan;
156 	nn->nfsd_reply_cache_shrinker.count_objects = nfsd_reply_cache_count;
157 	nn->nfsd_reply_cache_shrinker.seeks = 1;
158 	status = register_shrinker(&nn->nfsd_reply_cache_shrinker);
159 	if (status)
160 		goto out_nomem;
161 
162 	nn->drc_slab = kmem_cache_create("nfsd_drc",
163 				sizeof(struct svc_cacherep), 0, 0, NULL);
164 	if (!nn->drc_slab)
165 		goto out_shrinker;
166 
167 	nn->drc_hashtbl = kcalloc(hashsize,
168 				sizeof(*nn->drc_hashtbl), GFP_KERNEL);
169 	if (!nn->drc_hashtbl) {
170 		nn->drc_hashtbl = vzalloc(array_size(hashsize,
171 						 sizeof(*nn->drc_hashtbl)));
172 		if (!nn->drc_hashtbl)
173 			goto out_slab;
174 	}
175 
176 	for (i = 0; i < hashsize; i++) {
177 		INIT_LIST_HEAD(&nn->drc_hashtbl[i].lru_head);
178 		spin_lock_init(&nn->drc_hashtbl[i].cache_lock);
179 	}
180 	nn->drc_hashsize = hashsize;
181 
182 	return 0;
183 out_slab:
184 	kmem_cache_destroy(nn->drc_slab);
185 out_shrinker:
186 	unregister_shrinker(&nn->nfsd_reply_cache_shrinker);
187 out_nomem:
188 	printk(KERN_ERR "nfsd: failed to allocate reply cache\n");
189 	return -ENOMEM;
190 }
191 
192 void nfsd_reply_cache_shutdown(struct nfsd_net *nn)
193 {
194 	struct svc_cacherep	*rp;
195 	unsigned int i;
196 
197 	unregister_shrinker(&nn->nfsd_reply_cache_shrinker);
198 
199 	for (i = 0; i < nn->drc_hashsize; i++) {
200 		struct list_head *head = &nn->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(&nn->drc_hashtbl[i],
204 									rp, nn);
205 		}
206 	}
207 
208 	kvfree(nn->drc_hashtbl);
209 	nn->drc_hashtbl = NULL;
210 	nn->drc_hashsize = 0;
211 
212 	kmem_cache_destroy(nn->drc_slab);
213 	nn->drc_slab = NULL;
214 }
215 
216 /*
217  * Move cache entry to end of LRU list, and queue the cleaner to run if it's
218  * not already scheduled.
219  */
220 static void
221 lru_put_end(struct nfsd_drc_bucket *b, struct svc_cacherep *rp)
222 {
223 	rp->c_timestamp = jiffies;
224 	list_move_tail(&rp->c_lru, &b->lru_head);
225 }
226 
227 static long
228 prune_bucket(struct nfsd_drc_bucket *b, struct nfsd_net *nn)
229 {
230 	struct svc_cacherep *rp, *tmp;
231 	long freed = 0;
232 
233 	list_for_each_entry_safe(rp, tmp, &b->lru_head, c_lru) {
234 		/*
235 		 * Don't free entries attached to calls that are still
236 		 * in-progress, but do keep scanning the list.
237 		 */
238 		if (rp->c_state == RC_INPROG)
239 			continue;
240 		if (atomic_read(&nn->num_drc_entries) <= nn->max_drc_entries &&
241 		    time_before(jiffies, rp->c_timestamp + RC_EXPIRE))
242 			break;
243 		nfsd_reply_cache_free_locked(b, rp, nn);
244 		freed++;
245 	}
246 	return freed;
247 }
248 
249 /*
250  * Walk the LRU list and prune off entries that are older than RC_EXPIRE.
251  * Also prune the oldest ones when the total exceeds the max number of entries.
252  */
253 static long
254 prune_cache_entries(struct nfsd_net *nn)
255 {
256 	unsigned int i;
257 	long freed = 0;
258 
259 	for (i = 0; i < nn->drc_hashsize; i++) {
260 		struct nfsd_drc_bucket *b = &nn->drc_hashtbl[i];
261 
262 		if (list_empty(&b->lru_head))
263 			continue;
264 		spin_lock(&b->cache_lock);
265 		freed += prune_bucket(b, nn);
266 		spin_unlock(&b->cache_lock);
267 	}
268 	return freed;
269 }
270 
271 static unsigned long
272 nfsd_reply_cache_count(struct shrinker *shrink, struct shrink_control *sc)
273 {
274 	struct nfsd_net *nn = container_of(shrink,
275 				struct nfsd_net, nfsd_reply_cache_shrinker);
276 
277 	return atomic_read(&nn->num_drc_entries);
278 }
279 
280 static unsigned long
281 nfsd_reply_cache_scan(struct shrinker *shrink, struct shrink_control *sc)
282 {
283 	struct nfsd_net *nn = container_of(shrink,
284 				struct nfsd_net, nfsd_reply_cache_shrinker);
285 
286 	return prune_cache_entries(nn);
287 }
288 /*
289  * Walk an xdr_buf and get a CRC for at most the first RC_CSUMLEN bytes
290  */
291 static __wsum
292 nfsd_cache_csum(struct svc_rqst *rqstp)
293 {
294 	int idx;
295 	unsigned int base;
296 	__wsum csum;
297 	struct xdr_buf *buf = &rqstp->rq_arg;
298 	const unsigned char *p = buf->head[0].iov_base;
299 	size_t csum_len = min_t(size_t, buf->head[0].iov_len + buf->page_len,
300 				RC_CSUMLEN);
301 	size_t len = min(buf->head[0].iov_len, csum_len);
302 
303 	/* rq_arg.head first */
304 	csum = csum_partial(p, len, 0);
305 	csum_len -= len;
306 
307 	/* Continue into page array */
308 	idx = buf->page_base / PAGE_SIZE;
309 	base = buf->page_base & ~PAGE_MASK;
310 	while (csum_len) {
311 		p = page_address(buf->pages[idx]) + base;
312 		len = min_t(size_t, PAGE_SIZE - base, csum_len);
313 		csum = csum_partial(p, len, csum);
314 		csum_len -= len;
315 		base = 0;
316 		++idx;
317 	}
318 	return csum;
319 }
320 
321 static int
322 nfsd_cache_key_cmp(const struct svc_cacherep *key,
323 			const struct svc_cacherep *rp, struct nfsd_net *nn)
324 {
325 	if (key->c_key.k_xid == rp->c_key.k_xid &&
326 	    key->c_key.k_csum != rp->c_key.k_csum)
327 		++nn->payload_misses;
328 
329 	return memcmp(&key->c_key, &rp->c_key, sizeof(key->c_key));
330 }
331 
332 /*
333  * Search the request hash for an entry that matches the given rqstp.
334  * Must be called with cache_lock held. Returns the found entry or
335  * inserts an empty key on failure.
336  */
337 static struct svc_cacherep *
338 nfsd_cache_insert(struct nfsd_drc_bucket *b, struct svc_cacherep *key,
339 			struct nfsd_net *nn)
340 {
341 	struct svc_cacherep	*rp, *ret = key;
342 	struct rb_node		**p = &b->rb_head.rb_node,
343 				*parent = NULL;
344 	unsigned int		entries = 0;
345 	int cmp;
346 
347 	while (*p != NULL) {
348 		++entries;
349 		parent = *p;
350 		rp = rb_entry(parent, struct svc_cacherep, c_node);
351 
352 		cmp = nfsd_cache_key_cmp(key, rp, nn);
353 		if (cmp < 0)
354 			p = &parent->rb_left;
355 		else if (cmp > 0)
356 			p = &parent->rb_right;
357 		else {
358 			ret = rp;
359 			goto out;
360 		}
361 	}
362 	rb_link_node(&key->c_node, parent, p);
363 	rb_insert_color(&key->c_node, &b->rb_head);
364 out:
365 	/* tally hash chain length stats */
366 	if (entries > nn->longest_chain) {
367 		nn->longest_chain = entries;
368 		nn->longest_chain_cachesize = atomic_read(&nn->num_drc_entries);
369 	} else if (entries == nn->longest_chain) {
370 		/* prefer to keep the smallest cachesize possible here */
371 		nn->longest_chain_cachesize = min_t(unsigned int,
372 				nn->longest_chain_cachesize,
373 				atomic_read(&nn->num_drc_entries));
374 	}
375 
376 	lru_put_end(b, ret);
377 	return ret;
378 }
379 
380 /*
381  * Try to find an entry matching the current call in the cache. When none
382  * is found, we try to grab the oldest expired entry off the LRU list. If
383  * a suitable one isn't there, then drop the cache_lock and allocate a
384  * new one, then search again in case one got inserted while this thread
385  * didn't hold the lock.
386  */
387 int
388 nfsd_cache_lookup(struct svc_rqst *rqstp)
389 {
390 	struct nfsd_net *nn = net_generic(SVC_NET(rqstp), nfsd_net_id);
391 	struct svc_cacherep	*rp, *found;
392 	__be32			xid = rqstp->rq_xid;
393 	__wsum			csum;
394 	u32 hash = nfsd_cache_hash(xid, nn);
395 	struct nfsd_drc_bucket *b = &nn->drc_hashtbl[hash];
396 	int type = rqstp->rq_cachetype;
397 	int rtn = RC_DOIT;
398 
399 	rqstp->rq_cacherep = NULL;
400 	if (type == RC_NOCACHE) {
401 		nfsdstats.rcnocache++;
402 		return rtn;
403 	}
404 
405 	csum = nfsd_cache_csum(rqstp);
406 
407 	/*
408 	 * Since the common case is a cache miss followed by an insert,
409 	 * preallocate an entry.
410 	 */
411 	rp = nfsd_reply_cache_alloc(rqstp, csum, nn);
412 	if (!rp) {
413 		dprintk("nfsd: unable to allocate DRC entry!\n");
414 		return rtn;
415 	}
416 
417 	spin_lock(&b->cache_lock);
418 	found = nfsd_cache_insert(b, rp, nn);
419 	if (found != rp) {
420 		nfsd_reply_cache_free_locked(NULL, rp, nn);
421 		rp = found;
422 		goto found_entry;
423 	}
424 
425 	nfsdstats.rcmisses++;
426 	rqstp->rq_cacherep = rp;
427 	rp->c_state = RC_INPROG;
428 
429 	atomic_inc(&nn->num_drc_entries);
430 	nn->drc_mem_usage += sizeof(*rp);
431 
432 	/* go ahead and prune the cache */
433 	prune_bucket(b, nn);
434  out:
435 	spin_unlock(&b->cache_lock);
436 	return rtn;
437 
438 found_entry:
439 	/* We found a matching entry which is either in progress or done. */
440 	nfsdstats.rchits++;
441 	rtn = RC_DROPIT;
442 
443 	/* Request being processed */
444 	if (rp->c_state == RC_INPROG)
445 		goto out;
446 
447 	/* From the hall of fame of impractical attacks:
448 	 * Is this a user who tries to snoop on the cache? */
449 	rtn = RC_DOIT;
450 	if (!test_bit(RQ_SECURE, &rqstp->rq_flags) && rp->c_secure)
451 		goto out;
452 
453 	/* Compose RPC reply header */
454 	switch (rp->c_type) {
455 	case RC_NOCACHE:
456 		break;
457 	case RC_REPLSTAT:
458 		svc_putu32(&rqstp->rq_res.head[0], rp->c_replstat);
459 		rtn = RC_REPLY;
460 		break;
461 	case RC_REPLBUFF:
462 		if (!nfsd_cache_append(rqstp, &rp->c_replvec))
463 			goto out;	/* should not happen */
464 		rtn = RC_REPLY;
465 		break;
466 	default:
467 		printk(KERN_WARNING "nfsd: bad repcache type %d\n", rp->c_type);
468 		nfsd_reply_cache_free_locked(b, rp, nn);
469 	}
470 
471 	goto out;
472 }
473 
474 /*
475  * Update a cache entry. This is called from nfsd_dispatch when
476  * the procedure has been executed and the complete reply is in
477  * rqstp->rq_res.
478  *
479  * We're copying around data here rather than swapping buffers because
480  * the toplevel loop requires max-sized buffers, which would be a waste
481  * of memory for a cache with a max reply size of 100 bytes (diropokres).
482  *
483  * If we should start to use different types of cache entries tailored
484  * specifically for attrstat and fh's, we may save even more space.
485  *
486  * Also note that a cachetype of RC_NOCACHE can legally be passed when
487  * nfsd failed to encode a reply that otherwise would have been cached.
488  * In this case, nfsd_cache_update is called with statp == NULL.
489  */
490 void
491 nfsd_cache_update(struct svc_rqst *rqstp, int cachetype, __be32 *statp)
492 {
493 	struct nfsd_net *nn = net_generic(SVC_NET(rqstp), nfsd_net_id);
494 	struct svc_cacherep *rp = rqstp->rq_cacherep;
495 	struct kvec	*resv = &rqstp->rq_res.head[0], *cachv;
496 	u32		hash;
497 	struct nfsd_drc_bucket *b;
498 	int		len;
499 	size_t		bufsize = 0;
500 
501 	if (!rp)
502 		return;
503 
504 	hash = nfsd_cache_hash(rp->c_key.k_xid, nn);
505 	b = &nn->drc_hashtbl[hash];
506 
507 	len = resv->iov_len - ((char*)statp - (char*)resv->iov_base);
508 	len >>= 2;
509 
510 	/* Don't cache excessive amounts of data and XDR failures */
511 	if (!statp || len > (256 >> 2)) {
512 		nfsd_reply_cache_free(b, rp, nn);
513 		return;
514 	}
515 
516 	switch (cachetype) {
517 	case RC_REPLSTAT:
518 		if (len != 1)
519 			printk("nfsd: RC_REPLSTAT/reply len %d!\n",len);
520 		rp->c_replstat = *statp;
521 		break;
522 	case RC_REPLBUFF:
523 		cachv = &rp->c_replvec;
524 		bufsize = len << 2;
525 		cachv->iov_base = kmalloc(bufsize, GFP_KERNEL);
526 		if (!cachv->iov_base) {
527 			nfsd_reply_cache_free(b, rp, nn);
528 			return;
529 		}
530 		cachv->iov_len = bufsize;
531 		memcpy(cachv->iov_base, statp, bufsize);
532 		break;
533 	case RC_NOCACHE:
534 		nfsd_reply_cache_free(b, rp, nn);
535 		return;
536 	}
537 	spin_lock(&b->cache_lock);
538 	nn->drc_mem_usage += bufsize;
539 	lru_put_end(b, rp);
540 	rp->c_secure = test_bit(RQ_SECURE, &rqstp->rq_flags);
541 	rp->c_type = cachetype;
542 	rp->c_state = RC_DONE;
543 	spin_unlock(&b->cache_lock);
544 	return;
545 }
546 
547 /*
548  * Copy cached reply to current reply buffer. Should always fit.
549  * FIXME as reply is in a page, we should just attach the page, and
550  * keep a refcount....
551  */
552 static int
553 nfsd_cache_append(struct svc_rqst *rqstp, struct kvec *data)
554 {
555 	struct kvec	*vec = &rqstp->rq_res.head[0];
556 
557 	if (vec->iov_len + data->iov_len > PAGE_SIZE) {
558 		printk(KERN_WARNING "nfsd: cached reply too large (%zd).\n",
559 				data->iov_len);
560 		return 0;
561 	}
562 	memcpy((char*)vec->iov_base + vec->iov_len, data->iov_base, data->iov_len);
563 	vec->iov_len += data->iov_len;
564 	return 1;
565 }
566 
567 /*
568  * Note that fields may be added, removed or reordered in the future. Programs
569  * scraping this file for info should test the labels to ensure they're
570  * getting the correct field.
571  */
572 static int nfsd_reply_cache_stats_show(struct seq_file *m, void *v)
573 {
574 	struct nfsd_net *nn = m->private;
575 
576 	seq_printf(m, "max entries:           %u\n", nn->max_drc_entries);
577 	seq_printf(m, "num entries:           %u\n",
578 			atomic_read(&nn->num_drc_entries));
579 	seq_printf(m, "hash buckets:          %u\n", 1 << nn->maskbits);
580 	seq_printf(m, "mem usage:             %u\n", nn->drc_mem_usage);
581 	seq_printf(m, "cache hits:            %u\n", nfsdstats.rchits);
582 	seq_printf(m, "cache misses:          %u\n", nfsdstats.rcmisses);
583 	seq_printf(m, "not cached:            %u\n", nfsdstats.rcnocache);
584 	seq_printf(m, "payload misses:        %u\n", nn->payload_misses);
585 	seq_printf(m, "longest chain len:     %u\n", nn->longest_chain);
586 	seq_printf(m, "cachesize at longest:  %u\n", nn->longest_chain_cachesize);
587 	return 0;
588 }
589 
590 int nfsd_reply_cache_stats_open(struct inode *inode, struct file *file)
591 {
592 	struct nfsd_net *nn = net_generic(file_inode(file)->i_sb->s_fs_info,
593 								nfsd_net_id);
594 
595 	return single_open(file, nfsd_reply_cache_stats_show, nn);
596 }
597