xref: /illumos-gate/usr/src/uts/common/fs/dnlc.c (revision e7b9901e)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2015, Joyent, Inc.
24  * Copyright (c) 2017 by Delphix. All rights reserved.
25  */
26 
27 /*	Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T	*/
28 /*	  All Rights Reserved	*/
29 
30 /*
31  * University Copyright- Copyright (c) 1982, 1986, 1988
32  * The Regents of the University of California
33  * All Rights Reserved
34  *
35  * University Acknowledgment- Portions of this document are derived from
36  * software developed by the University of California, Berkeley, and its
37  * contributors.
38  */
39 
40 #include <sys/types.h>
41 #include <sys/systm.h>
42 #include <sys/param.h>
43 #include <sys/t_lock.h>
44 #include <sys/systm.h>
45 #include <sys/vfs.h>
46 #include <sys/vnode.h>
47 #include <sys/dnlc.h>
48 #include <sys/kmem.h>
49 #include <sys/cmn_err.h>
50 #include <sys/vtrace.h>
51 #include <sys/bitmap.h>
52 #include <sys/var.h>
53 #include <sys/sysmacros.h>
54 #include <sys/kstat.h>
55 #include <sys/atomic.h>
56 #include <sys/taskq.h>
57 
58 /*
59  * Directory name lookup cache.
60  * Based on code originally done by Robert Elz at Melbourne.
61  *
62  * Names found by directory scans are retained in a cache
63  * for future reference.  Each hash chain is ordered by LRU
64  * Cache is indexed by hash value obtained from (vp, name)
65  * where the vp refers to the directory containing the name.
66  */
67 
68 /*
69  * We want to be able to identify files that are referenced only by the DNLC.
70  * When adding a reference from the DNLC, call VN_HOLD_DNLC instead of VN_HOLD,
71  * since multiple DNLC references should only be counted once in v_count. The
72  * VN_HOLD macro itself is aliased to VN_HOLD_CALLER in this file to help
73  * differentiate the behaviors.  (Unfortunately it is not possible to #undef
74  * VN_HOLD and retain VN_HOLD_CALLER. Ideally a Makefile rule would grep
75  * uncommented C tokens to check that VN_HOLD is referenced only once in this
76  * file, to define VN_HOLD_CALLER.)
77  */
78 #define	VN_HOLD_CALLER	VN_HOLD
79 #define	VN_HOLD_DNLC(vp)	{	\
80 	mutex_enter(&(vp)->v_lock);	\
81 	if ((vp)->v_count_dnlc == 0) {	\
82 		VN_HOLD_LOCKED(vp);	\
83 	}				\
84 	(vp)->v_count_dnlc++;		\
85 	mutex_exit(&(vp)->v_lock);	\
86 }
87 #define	VN_RELE_DNLC(vp)	{	\
88 	vn_rele_dnlc(vp);		\
89 }
90 
91 /*
92  * Tunable nc_hashavelen is the average length desired for this chain, from
93  * which the size of the nc_hash table is derived at create time.
94  */
95 #define	NC_HASHAVELEN_DEFAULT	4
96 int nc_hashavelen = NC_HASHAVELEN_DEFAULT;
97 
98 /*
99  * NC_MOVETOFRONT is the move-to-front threshold: if the hash lookup
100  * depth exceeds this value, we move the looked-up entry to the front of
101  * its hash chain.  The idea is to make sure that the most frequently
102  * accessed entries are found most quickly (by keeping them near the
103  * front of their hash chains).
104  */
105 #define	NC_MOVETOFRONT	2
106 
107 /*
108  *
109  * DNLC_MAX_RELE is used to size an array on the stack when releasing
110  * vnodes. This array is used rather than calling VN_RELE() inline because
111  * all dnlc locks must be dropped by that time in order to avoid a
112  * possible deadlock. This deadlock occurs when the dnlc holds the last
113  * reference to the vnode and so the VOP_INACTIVE vector is called which
114  * can in turn call back into the dnlc. A global array was used but had
115  * many problems:
116  *	1) Actually doesn't have an upper bound on the array size as
117  *	   entries can be added after starting the purge.
118  *	2) The locking scheme causes a hang.
119  *	3) Caused serialisation on the global lock.
120  *	4) The array was often unnecessarily huge.
121  *
122  * Note the current value 8 allows up to 4 cache entries (to be purged
123  * from each hash chain), before having to cycle around and retry.
124  * This ought to be ample given that nc_hashavelen is typically very small.
125  */
126 #define	DNLC_MAX_RELE	8 /* must be even */
127 
128 /*
129  * Hash table of name cache entries for fast lookup, dynamically
130  * allocated at startup.
131  */
132 nc_hash_t *nc_hash;
133 
134 /*
135  * Rotors. Used to select entries on a round-robin basis.
136  */
137 static nc_hash_t *dnlc_purge_fs1_rotor;
138 static nc_hash_t *dnlc_free_rotor;
139 
140 /*
141  * # of dnlc entries (uninitialized)
142  *
143  * the initial value was chosen as being
144  * a random string of bits, probably not
145  * normally chosen by a systems administrator
146  */
147 int ncsize = -1;
148 volatile uint32_t dnlc_nentries = 0;	/* current num of name cache entries */
149 static int nc_hashsz;			/* size of hash table */
150 static int nc_hashmask;			/* size of hash table minus 1 */
151 
152 /*
153  * The dnlc_reduce_cache() taskq queue is activated when there are
154  * ncsize name cache entries and if no parameter is provided, it reduces
155  * the size down to dnlc_nentries_low_water, which is by default one
156  * hundreth less (or 99%) of ncsize.
157  *
158  * If a parameter is provided to dnlc_reduce_cache(), then we reduce
159  * the size down based on ncsize_onepercent - where ncsize_onepercent
160  * is 1% of ncsize; however, we never let dnlc_reduce_cache() reduce
161  * the size below 3% of ncsize (ncsize_min_percent).
162  */
163 #define	DNLC_LOW_WATER_DIVISOR_DEFAULT 100
164 uint_t dnlc_low_water_divisor = DNLC_LOW_WATER_DIVISOR_DEFAULT;
165 uint_t dnlc_nentries_low_water;
166 int dnlc_reduce_idle = 1; /* no locking needed */
167 uint_t ncsize_onepercent;
168 uint_t ncsize_min_percent;
169 
170 /*
171  * If dnlc_nentries hits dnlc_max_nentries (twice ncsize)
172  * then this means the dnlc_reduce_cache() taskq is failing to
173  * keep up. In this case we refuse to add new entries to the dnlc
174  * until the taskq catches up.
175  */
176 uint_t dnlc_max_nentries; /* twice ncsize */
177 uint64_t dnlc_max_nentries_cnt = 0; /* statistic on times we failed */
178 
179 /*
180  * Tunable to define when we should just remove items from
181  * the end of the chain.
182  */
183 #define	DNLC_LONG_CHAIN 8
184 uint_t dnlc_long_chain = DNLC_LONG_CHAIN;
185 
186 /*
187  * ncstats has been deprecated, due to the integer size of the counters
188  * which can easily overflow in the dnlc.
189  * It is maintained (at some expense) for compatability.
190  * The preferred interface is the kstat accessible nc_stats below.
191  */
192 struct ncstats ncstats;
193 
194 struct nc_stats ncs = {
195 	{ "hits",			KSTAT_DATA_UINT64 },
196 	{ "misses",			KSTAT_DATA_UINT64 },
197 	{ "negative_cache_hits",	KSTAT_DATA_UINT64 },
198 	{ "enters",			KSTAT_DATA_UINT64 },
199 	{ "double_enters",		KSTAT_DATA_UINT64 },
200 	{ "purge_total_entries",	KSTAT_DATA_UINT64 },
201 	{ "purge_all",			KSTAT_DATA_UINT64 },
202 	{ "purge_vp",			KSTAT_DATA_UINT64 },
203 	{ "purge_vfs",			KSTAT_DATA_UINT64 },
204 	{ "purge_fs1",			KSTAT_DATA_UINT64 },
205 	{ "pick_free",			KSTAT_DATA_UINT64 },
206 	{ "pick_heuristic",		KSTAT_DATA_UINT64 },
207 	{ "pick_last",			KSTAT_DATA_UINT64 },
208 
209 	/* directory caching stats */
210 
211 	{ "dir_hits",			KSTAT_DATA_UINT64 },
212 	{ "dir_misses",			KSTAT_DATA_UINT64 },
213 	{ "dir_cached_current",		KSTAT_DATA_UINT64 },
214 	{ "dir_entries_cached_current",	KSTAT_DATA_UINT64 },
215 	{ "dir_cached_total",		KSTAT_DATA_UINT64 },
216 	{ "dir_start_no_memory",	KSTAT_DATA_UINT64 },
217 	{ "dir_add_no_memory",		KSTAT_DATA_UINT64 },
218 	{ "dir_add_abort",		KSTAT_DATA_UINT64 },
219 	{ "dir_add_max",		KSTAT_DATA_UINT64 },
220 	{ "dir_remove_entry_fail",	KSTAT_DATA_UINT64 },
221 	{ "dir_remove_space_fail",	KSTAT_DATA_UINT64 },
222 	{ "dir_update_fail",		KSTAT_DATA_UINT64 },
223 	{ "dir_fini_purge",		KSTAT_DATA_UINT64 },
224 	{ "dir_reclaim_last",		KSTAT_DATA_UINT64 },
225 	{ "dir_reclaim_any",		KSTAT_DATA_UINT64 },
226 };
227 
228 static int doingcache = 1;
229 
230 vnode_t negative_cache_vnode;
231 
232 /*
233  * Insert entry at the front of the queue
234  */
235 #define	nc_inshash(ncp, hp) \
236 { \
237 	(ncp)->hash_next = (hp)->hash_next; \
238 	(ncp)->hash_prev = (ncache_t *)(hp); \
239 	(hp)->hash_next->hash_prev = (ncp); \
240 	(hp)->hash_next = (ncp); \
241 }
242 
243 /*
244  * Remove entry from hash queue
245  */
246 #define	nc_rmhash(ncp) \
247 { \
248 	(ncp)->hash_prev->hash_next = (ncp)->hash_next; \
249 	(ncp)->hash_next->hash_prev = (ncp)->hash_prev; \
250 	(ncp)->hash_prev = NULL; \
251 	(ncp)->hash_next = NULL; \
252 }
253 
254 /*
255  * Free an entry.
256  */
257 #define	dnlc_free(ncp) \
258 { \
259 	kmem_free((ncp), NCACHE_SIZE((ncp)->namlen)); \
260 	atomic_dec_32(&dnlc_nentries); \
261 }
262 
263 
264 /*
265  * Cached directory info.
266  * ======================
267  */
268 
269 /*
270  * Cached directory free space hash function.
271  * Needs the free space handle and the dcp to get the hash table size
272  * Returns the hash index.
273  */
274 #define	DDFHASH(handle, dcp) ((handle >> 2) & (dcp)->dc_fhash_mask)
275 
276 /*
277  * Cached directory name entry hash function.
278  * Uses the name and returns in the input arguments the hash and the name
279  * length.
280  */
281 #define	DNLC_DIR_HASH(name, hash, namelen)			\
282 	{							\
283 		char Xc;					\
284 		const char *Xcp;				\
285 		hash = *name;					\
286 		for (Xcp = (name + 1); (Xc = *Xcp) != 0; Xcp++)	\
287 			hash = (hash << 4) + hash + Xc;		\
288 		ASSERT((Xcp - (name)) <= ((1 << NBBY) - 1));	\
289 		namelen = Xcp - (name);				\
290 	}
291 
292 /* special dircache_t pointer to indicate error should be returned */
293 /*
294  * The anchor directory cache pointer can contain 3 types of values,
295  * 1) NULL: No directory cache
296  * 2) DC_RET_LOW_MEM (-1): There was a directory cache that found to be
297  *    too big or a memory shortage occurred. This value remains in the
298  *    pointer until a dnlc_dir_start() which returns the a DNOMEM error.
299  *    This is kludgy but efficient and only visible in this source file.
300  * 3) A valid cache pointer.
301  */
302 #define	DC_RET_LOW_MEM (dircache_t *)1
303 #define	VALID_DIR_CACHE(dcp) ((dircache_t *)(dcp) > DC_RET_LOW_MEM)
304 
305 /* Tunables */
306 uint_t dnlc_dir_enable = 1; /* disable caching directories by setting to 0 */
307 uint_t dnlc_dir_min_size = 40; /* min no of directory entries before caching */
308 uint_t dnlc_dir_max_size = UINT_MAX; /* ditto maximum */
309 uint_t dnlc_dir_hash_size_shift = 3; /* 8 entries per hash bucket */
310 uint_t dnlc_dir_min_reclaim =  350000; /* approx 1MB of dcentrys */
311 /*
312  * dnlc_dir_hash_resize_shift determines when the hash tables
313  * get re-adjusted due to growth or shrinkage
314  * - currently 2 indicating that there can be at most 4
315  * times or at least one quarter the number of entries
316  * before hash table readjustment. Note that with
317  * dnlc_dir_hash_size_shift above set at 3 this would
318  * mean readjustment would occur if the average number
319  * of entries went above 32 or below 2
320  */
321 uint_t dnlc_dir_hash_resize_shift = 2; /* readjust rate */
322 
323 static kmem_cache_t *dnlc_dir_space_cache; /* free space entry cache */
324 static dchead_t dc_head; /* anchor of cached directories */
325 
326 /* Prototypes */
327 static ncache_t *dnlc_get(uchar_t namlen);
328 static ncache_t *dnlc_search(vnode_t *dp, const char *name, uchar_t namlen,
329     int hash);
330 static void dnlc_dir_reclaim(void *unused);
331 static void dnlc_dir_abort(dircache_t *dcp);
332 static void dnlc_dir_adjust_fhash(dircache_t *dcp);
333 static void dnlc_dir_adjust_nhash(dircache_t *dcp);
334 static void do_dnlc_reduce_cache(void *);
335 
336 
337 /*
338  * Initialize the directory cache.
339  */
340 void
dnlc_init()341 dnlc_init()
342 {
343 	nc_hash_t *hp;
344 	kstat_t *ksp;
345 	int i;
346 
347 	/*
348 	 * Set up the size of the dnlc (ncsize) and its low water mark.
349 	 */
350 	if (ncsize == -1) {
351 		/* calculate a reasonable size for the low water */
352 		dnlc_nentries_low_water = 4 * (v.v_proc + maxusers) + 320;
353 		ncsize = dnlc_nentries_low_water +
354 		    (dnlc_nentries_low_water / dnlc_low_water_divisor);
355 	} else {
356 		/* don't change the user specified ncsize */
357 		dnlc_nentries_low_water =
358 		    ncsize - (ncsize / dnlc_low_water_divisor);
359 	}
360 	if (ncsize <= 0) {
361 		doingcache = 0;
362 		dnlc_dir_enable = 0; /* also disable directory caching */
363 		ncsize = 0;
364 		cmn_err(CE_NOTE, "name cache (dnlc) disabled");
365 		return;
366 	}
367 	dnlc_max_nentries = ncsize * 2;
368 	ncsize_onepercent = ncsize / 100;
369 	ncsize_min_percent = ncsize_onepercent * 3;
370 
371 	/*
372 	 * Initialise the hash table.
373 	 * Compute hash size rounding to the next power of two.
374 	 */
375 	nc_hashsz = ncsize / nc_hashavelen;
376 	nc_hashsz = 1 << highbit(nc_hashsz);
377 	nc_hashmask = nc_hashsz - 1;
378 	nc_hash = kmem_zalloc(nc_hashsz * sizeof (*nc_hash), KM_SLEEP);
379 	for (i = 0; i < nc_hashsz; i++) {
380 		hp = (nc_hash_t *)&nc_hash[i];
381 		mutex_init(&hp->hash_lock, NULL, MUTEX_DEFAULT, NULL);
382 		hp->hash_next = (ncache_t *)hp;
383 		hp->hash_prev = (ncache_t *)hp;
384 	}
385 
386 	/*
387 	 * Initialize rotors
388 	 */
389 	dnlc_free_rotor = dnlc_purge_fs1_rotor = &nc_hash[0];
390 
391 	/*
392 	 * Set up the directory caching to use kmem_cache_alloc
393 	 * for its free space entries so that we can get a callback
394 	 * when the system is short on memory, to allow us to free
395 	 * up some memory. we don't use the constructor/deconstructor
396 	 * functions.
397 	 */
398 	dnlc_dir_space_cache = kmem_cache_create("dnlc_space_cache",
399 	    sizeof (dcfree_t), 0, NULL, NULL, dnlc_dir_reclaim, NULL,
400 	    NULL, 0);
401 
402 	/*
403 	 * Initialise the head of the cached directory structures
404 	 */
405 	mutex_init(&dc_head.dch_lock, NULL, MUTEX_DEFAULT, NULL);
406 	dc_head.dch_next = (dircache_t *)&dc_head;
407 	dc_head.dch_prev = (dircache_t *)&dc_head;
408 
409 	/*
410 	 * Put a hold on the negative cache vnode so that it never goes away
411 	 * (VOP_INACTIVE isn't called on it).
412 	 */
413 	vn_reinit(&negative_cache_vnode);
414 
415 	/*
416 	 * Initialise kstats - both the old compatability raw kind and
417 	 * the more extensive named stats.
418 	 */
419 	ksp = kstat_create("unix", 0, "ncstats", "misc", KSTAT_TYPE_RAW,
420 	    sizeof (struct ncstats), KSTAT_FLAG_VIRTUAL);
421 	if (ksp) {
422 		ksp->ks_data = (void *) &ncstats;
423 		kstat_install(ksp);
424 	}
425 	ksp = kstat_create("unix", 0, "dnlcstats", "misc", KSTAT_TYPE_NAMED,
426 	    sizeof (ncs) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
427 	if (ksp) {
428 		ksp->ks_data = (void *) &ncs;
429 		kstat_install(ksp);
430 	}
431 }
432 
433 /*
434  * Add a name to the directory cache.
435  */
436 void
dnlc_enter(vnode_t * dp,const char * name,vnode_t * vp)437 dnlc_enter(vnode_t *dp, const char *name, vnode_t *vp)
438 {
439 	ncache_t *ncp;
440 	nc_hash_t *hp;
441 	uchar_t namlen;
442 	int hash;
443 
444 	TRACE_0(TR_FAC_NFS, TR_DNLC_ENTER_START, "dnlc_enter_start:");
445 
446 	if (!doingcache) {
447 		TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
448 		    "dnlc_enter_end:(%S) %d", "not caching", 0);
449 		return;
450 	}
451 
452 	/*
453 	 * Get a new dnlc entry. Assume the entry won't be in the cache
454 	 * and initialize it now
455 	 */
456 	DNLCHASH(name, dp, hash, namlen);
457 	if ((ncp = dnlc_get(namlen)) == NULL)
458 		return;
459 	ncp->dp = dp;
460 	VN_HOLD_DNLC(dp);
461 	ncp->vp = vp;
462 	VN_HOLD_DNLC(vp);
463 	bcopy(name, ncp->name, namlen);
464 	ncp->hash = hash;
465 	hp = &nc_hash[hash & nc_hashmask];
466 
467 	mutex_enter(&hp->hash_lock);
468 	if (dnlc_search(dp, name, namlen, hash) != NULL) {
469 		mutex_exit(&hp->hash_lock);
470 		ncstats.dbl_enters++;
471 		ncs.ncs_dbl_enters.value.ui64++;
472 		VN_RELE_DNLC(dp);
473 		VN_RELE_DNLC(vp);
474 		dnlc_free(ncp);		/* crfree done here */
475 		TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
476 		    "dnlc_enter_end:(%S) %d", "dbl enter", ncstats.dbl_enters);
477 		return;
478 	}
479 	/*
480 	 * Insert back into the hash chain.
481 	 */
482 	nc_inshash(ncp, hp);
483 	mutex_exit(&hp->hash_lock);
484 	ncstats.enters++;
485 	ncs.ncs_enters.value.ui64++;
486 	TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
487 	    "dnlc_enter_end:(%S) %d", "done", ncstats.enters);
488 }
489 
490 /*
491  * Add a name to the directory cache.
492  *
493  * This function is basically identical with
494  * dnlc_enter().  The difference is that when the
495  * desired dnlc entry is found, the vnode in the
496  * ncache is compared with the vnode passed in.
497  *
498  * If they are not equal then the ncache is
499  * updated with the passed in vnode.  Otherwise
500  * it just frees up the newly allocated dnlc entry.
501  */
502 void
dnlc_update(vnode_t * dp,const char * name,vnode_t * vp)503 dnlc_update(vnode_t *dp, const char *name, vnode_t *vp)
504 {
505 	ncache_t *ncp;
506 	ncache_t *tcp;
507 	vnode_t *tvp;
508 	nc_hash_t *hp;
509 	int hash;
510 	uchar_t namlen;
511 
512 	TRACE_0(TR_FAC_NFS, TR_DNLC_ENTER_START, "dnlc_update_start:");
513 
514 	if (!doingcache) {
515 		TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
516 		    "dnlc_update_end:(%S) %d", "not caching", 0);
517 		return;
518 	}
519 
520 	/*
521 	 * Get a new dnlc entry and initialize it now.
522 	 * If we fail to get a new entry, call dnlc_remove() to purge
523 	 * any existing dnlc entry including negative cache (DNLC_NO_VNODE)
524 	 * entry.
525 	 * Failure to clear an existing entry could result in false dnlc
526 	 * lookup (negative/stale entry).
527 	 */
528 	DNLCHASH(name, dp, hash, namlen);
529 	if ((ncp = dnlc_get(namlen)) == NULL) {
530 		dnlc_remove(dp, name);
531 		return;
532 	}
533 	ncp->dp = dp;
534 	VN_HOLD_DNLC(dp);
535 	ncp->vp = vp;
536 	VN_HOLD_DNLC(vp);
537 	bcopy(name, ncp->name, namlen);
538 	ncp->hash = hash;
539 	hp = &nc_hash[hash & nc_hashmask];
540 
541 	mutex_enter(&hp->hash_lock);
542 	if ((tcp = dnlc_search(dp, name, namlen, hash)) != NULL) {
543 		if (tcp->vp != vp) {
544 			tvp = tcp->vp;
545 			tcp->vp = vp;
546 			mutex_exit(&hp->hash_lock);
547 			VN_RELE_DNLC(tvp);
548 			ncstats.enters++;
549 			ncs.ncs_enters.value.ui64++;
550 			TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
551 			    "dnlc_update_end:(%S) %d", "done", ncstats.enters);
552 		} else {
553 			mutex_exit(&hp->hash_lock);
554 			VN_RELE_DNLC(vp);
555 			ncstats.dbl_enters++;
556 			ncs.ncs_dbl_enters.value.ui64++;
557 			TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
558 			    "dnlc_update_end:(%S) %d",
559 			    "dbl enter", ncstats.dbl_enters);
560 		}
561 		VN_RELE_DNLC(dp);
562 		dnlc_free(ncp);		/* crfree done here */
563 		return;
564 	}
565 	/*
566 	 * insert the new entry, since it is not in dnlc yet
567 	 */
568 	nc_inshash(ncp, hp);
569 	mutex_exit(&hp->hash_lock);
570 	ncstats.enters++;
571 	ncs.ncs_enters.value.ui64++;
572 	TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
573 	    "dnlc_update_end:(%S) %d", "done", ncstats.enters);
574 }
575 
576 /*
577  * Look up a name in the directory name cache.
578  *
579  * Return a doubly-held vnode if found: one hold so that it may
580  * remain in the cache for other users, the other hold so that
581  * the cache is not re-cycled and the identity of the vnode is
582  * lost before the caller can use the vnode.
583  */
584 vnode_t *
dnlc_lookup(vnode_t * dp,const char * name)585 dnlc_lookup(vnode_t *dp, const char *name)
586 {
587 	ncache_t *ncp;
588 	nc_hash_t *hp;
589 	vnode_t *vp;
590 	int hash, depth;
591 	uchar_t namlen;
592 
593 	TRACE_2(TR_FAC_NFS, TR_DNLC_LOOKUP_START,
594 	    "dnlc_lookup_start:dp %x name %s", dp, name);
595 
596 	if (!doingcache) {
597 		TRACE_4(TR_FAC_NFS, TR_DNLC_LOOKUP_END,
598 		    "dnlc_lookup_end:%S %d vp %x name %s",
599 		    "not_caching", 0, NULL, name);
600 		return (NULL);
601 	}
602 
603 	DNLCHASH(name, dp, hash, namlen);
604 	depth = 1;
605 	hp = &nc_hash[hash & nc_hashmask];
606 	mutex_enter(&hp->hash_lock);
607 
608 	for (ncp = hp->hash_next; ncp != (ncache_t *)hp;
609 	    ncp = ncp->hash_next) {
610 		if (ncp->hash == hash &&	/* fast signature check */
611 		    ncp->dp == dp &&
612 		    ncp->namlen == namlen &&
613 		    bcmp(ncp->name, name, namlen) == 0) {
614 			/*
615 			 * Move this entry to the head of its hash chain
616 			 * if it's not already close.
617 			 */
618 			if (depth > NC_MOVETOFRONT) {
619 				ncache_t *next = ncp->hash_next;
620 				ncache_t *prev = ncp->hash_prev;
621 
622 				prev->hash_next = next;
623 				next->hash_prev = prev;
624 				ncp->hash_next = next = hp->hash_next;
625 				ncp->hash_prev = (ncache_t *)hp;
626 				next->hash_prev = ncp;
627 				hp->hash_next = ncp;
628 
629 				ncstats.move_to_front++;
630 			}
631 
632 			/*
633 			 * Put a hold on the vnode now so its identity
634 			 * can't change before the caller has a chance to
635 			 * put a hold on it.
636 			 */
637 			vp = ncp->vp;
638 			VN_HOLD_CALLER(vp);
639 			mutex_exit(&hp->hash_lock);
640 			ncstats.hits++;
641 			ncs.ncs_hits.value.ui64++;
642 			if (vp == DNLC_NO_VNODE) {
643 				ncs.ncs_neg_hits.value.ui64++;
644 			}
645 			TRACE_4(TR_FAC_NFS, TR_DNLC_LOOKUP_END,
646 			    "dnlc_lookup_end:%S %d vp %x name %s", "hit",
647 			    ncstats.hits, vp, name);
648 			return (vp);
649 		}
650 		depth++;
651 	}
652 
653 	mutex_exit(&hp->hash_lock);
654 	ncstats.misses++;
655 	ncs.ncs_misses.value.ui64++;
656 	TRACE_4(TR_FAC_NFS, TR_DNLC_LOOKUP_END,
657 	    "dnlc_lookup_end:%S %d vp %x name %s", "miss", ncstats.misses,
658 	    NULL, name);
659 	return (NULL);
660 }
661 
662 /*
663  * Remove an entry in the directory name cache.
664  */
665 void
dnlc_remove(vnode_t * dp,const char * name)666 dnlc_remove(vnode_t *dp, const char *name)
667 {
668 	ncache_t *ncp;
669 	nc_hash_t *hp;
670 	uchar_t namlen;
671 	int hash;
672 
673 	if (!doingcache)
674 		return;
675 	DNLCHASH(name, dp, hash, namlen);
676 	hp = &nc_hash[hash & nc_hashmask];
677 
678 	mutex_enter(&hp->hash_lock);
679 	if (ncp = dnlc_search(dp, name, namlen, hash)) {
680 		/*
681 		 * Free up the entry
682 		 */
683 		nc_rmhash(ncp);
684 		mutex_exit(&hp->hash_lock);
685 		VN_RELE_DNLC(ncp->vp);
686 		VN_RELE_DNLC(ncp->dp);
687 		dnlc_free(ncp);
688 		return;
689 	}
690 	mutex_exit(&hp->hash_lock);
691 }
692 
693 /*
694  * Purge the entire cache.
695  */
696 void
dnlc_purge()697 dnlc_purge()
698 {
699 	nc_hash_t *nch;
700 	ncache_t *ncp;
701 	int index;
702 	int i;
703 	vnode_t *nc_rele[DNLC_MAX_RELE];
704 
705 	if (!doingcache)
706 		return;
707 
708 	ncstats.purges++;
709 	ncs.ncs_purge_all.value.ui64++;
710 
711 	for (nch = nc_hash; nch < &nc_hash[nc_hashsz]; nch++) {
712 		index = 0;
713 		mutex_enter(&nch->hash_lock);
714 		ncp = nch->hash_next;
715 		while (ncp != (ncache_t *)nch) {
716 			ncache_t *np;
717 
718 			np = ncp->hash_next;
719 			nc_rele[index++] = ncp->vp;
720 			nc_rele[index++] = ncp->dp;
721 
722 			nc_rmhash(ncp);
723 			dnlc_free(ncp);
724 			ncp = np;
725 			ncs.ncs_purge_total.value.ui64++;
726 			if (index == DNLC_MAX_RELE)
727 				break;
728 		}
729 		mutex_exit(&nch->hash_lock);
730 
731 		/* Release holds on all the vnodes now that we have no locks */
732 		for (i = 0; i < index; i++) {
733 			VN_RELE_DNLC(nc_rele[i]);
734 		}
735 		if (ncp != (ncache_t *)nch) {
736 			nch--; /* Do current hash chain again */
737 		}
738 	}
739 }
740 
741 /*
742  * Purge any cache entries referencing a vnode. Exit as soon as the dnlc
743  * reference count goes to zero (the caller still holds a reference).
744  */
745 void
dnlc_purge_vp(vnode_t * vp)746 dnlc_purge_vp(vnode_t *vp)
747 {
748 	nc_hash_t *nch;
749 	ncache_t *ncp;
750 	int index;
751 	vnode_t *nc_rele[DNLC_MAX_RELE];
752 
753 	ASSERT(vp->v_count > 0);
754 	if (vp->v_count_dnlc == 0) {
755 		return;
756 	}
757 
758 	if (!doingcache)
759 		return;
760 
761 	ncstats.purges++;
762 	ncs.ncs_purge_vp.value.ui64++;
763 
764 	for (nch = nc_hash; nch < &nc_hash[nc_hashsz]; nch++) {
765 		index = 0;
766 		mutex_enter(&nch->hash_lock);
767 		ncp = nch->hash_next;
768 		while (ncp != (ncache_t *)nch) {
769 			ncache_t *np;
770 
771 			np = ncp->hash_next;
772 			if (ncp->dp == vp || ncp->vp == vp) {
773 				nc_rele[index++] = ncp->vp;
774 				nc_rele[index++] = ncp->dp;
775 				nc_rmhash(ncp);
776 				dnlc_free(ncp);
777 				ncs.ncs_purge_total.value.ui64++;
778 				if (index == DNLC_MAX_RELE) {
779 					ncp = np;
780 					break;
781 				}
782 			}
783 			ncp = np;
784 		}
785 		mutex_exit(&nch->hash_lock);
786 
787 		/* Release holds on all the vnodes now that we have no locks */
788 		while (index) {
789 			VN_RELE_DNLC(nc_rele[--index]);
790 		}
791 
792 		if (vp->v_count_dnlc == 0) {
793 			return;
794 		}
795 
796 		if (ncp != (ncache_t *)nch) {
797 			nch--; /* Do current hash chain again */
798 		}
799 	}
800 }
801 
802 /*
803  * Purge cache entries referencing a vfsp.  Caller supplies a count
804  * of entries to purge; up to that many will be freed.  A count of
805  * zero indicates that all such entries should be purged.  Returns
806  * the number of entries that were purged.
807  */
808 int
dnlc_purge_vfsp(vfs_t * vfsp,int count)809 dnlc_purge_vfsp(vfs_t *vfsp, int count)
810 {
811 	nc_hash_t *nch;
812 	ncache_t *ncp;
813 	int n = 0;
814 	int index;
815 	int i;
816 	vnode_t *nc_rele[DNLC_MAX_RELE];
817 
818 	if (!doingcache)
819 		return (0);
820 
821 	ncstats.purges++;
822 	ncs.ncs_purge_vfs.value.ui64++;
823 
824 	for (nch = nc_hash; nch < &nc_hash[nc_hashsz]; nch++) {
825 		index = 0;
826 		mutex_enter(&nch->hash_lock);
827 		ncp = nch->hash_next;
828 		while (ncp != (ncache_t *)nch) {
829 			ncache_t *np;
830 
831 			np = ncp->hash_next;
832 			ASSERT(ncp->dp != NULL);
833 			ASSERT(ncp->vp != NULL);
834 			if ((ncp->dp->v_vfsp == vfsp) ||
835 			    (ncp->vp->v_vfsp == vfsp)) {
836 				n++;
837 				nc_rele[index++] = ncp->vp;
838 				nc_rele[index++] = ncp->dp;
839 				nc_rmhash(ncp);
840 				dnlc_free(ncp);
841 				ncs.ncs_purge_total.value.ui64++;
842 				if (index == DNLC_MAX_RELE) {
843 					ncp = np;
844 					break;
845 				}
846 				if (count != 0 && n >= count) {
847 					break;
848 				}
849 			}
850 			ncp = np;
851 		}
852 		mutex_exit(&nch->hash_lock);
853 		/* Release holds on all the vnodes now that we have no locks */
854 		for (i = 0; i < index; i++) {
855 			VN_RELE_DNLC(nc_rele[i]);
856 		}
857 		if (count != 0 && n >= count) {
858 			return (n);
859 		}
860 		if (ncp != (ncache_t *)nch) {
861 			nch--; /* Do current hash chain again */
862 		}
863 	}
864 	return (n);
865 }
866 
867 /*
868  * Purge 1 entry from the dnlc that is part of the filesystem(s)
869  * represented by 'vop'. The purpose of this routine is to allow
870  * users of the dnlc to free a vnode that is being held by the dnlc.
871  *
872  * If we find a vnode that we release which will result in
873  * freeing the underlying vnode (count was 1), return 1, 0
874  * if no appropriate vnodes found.
875  *
876  * Note, vop is not the 'right' identifier for a filesystem.
877  */
878 int
dnlc_fs_purge1(vnodeops_t * vop)879 dnlc_fs_purge1(vnodeops_t *vop)
880 {
881 	nc_hash_t *end;
882 	nc_hash_t *hp;
883 	ncache_t *ncp;
884 	vnode_t *vp;
885 
886 	if (!doingcache)
887 		return (0);
888 
889 	ncs.ncs_purge_fs1.value.ui64++;
890 
891 	/*
892 	 * Scan the dnlc entries looking for a likely candidate.
893 	 */
894 	hp = end = dnlc_purge_fs1_rotor;
895 
896 	do {
897 		if (++hp == &nc_hash[nc_hashsz])
898 			hp = nc_hash;
899 		dnlc_purge_fs1_rotor = hp;
900 		if (hp->hash_next == (ncache_t *)hp)
901 			continue;
902 		mutex_enter(&hp->hash_lock);
903 		for (ncp = hp->hash_prev;
904 		    ncp != (ncache_t *)hp;
905 		    ncp = ncp->hash_prev) {
906 			vp = ncp->vp;
907 			if (!vn_has_cached_data(vp) && (vp->v_count == 1) &&
908 			    vn_matchops(vp, vop))
909 				break;
910 		}
911 		if (ncp != (ncache_t *)hp) {
912 			nc_rmhash(ncp);
913 			mutex_exit(&hp->hash_lock);
914 			VN_RELE_DNLC(ncp->dp);
915 			VN_RELE_DNLC(vp)
916 			dnlc_free(ncp);
917 			ncs.ncs_purge_total.value.ui64++;
918 			return (1);
919 		}
920 		mutex_exit(&hp->hash_lock);
921 	} while (hp != end);
922 	return (0);
923 }
924 
925 /*
926  * Utility routine to search for a cache entry. Return the
927  * ncache entry if found, NULL otherwise.
928  */
929 static ncache_t *
dnlc_search(vnode_t * dp,const char * name,uchar_t namlen,int hash)930 dnlc_search(vnode_t *dp, const char *name, uchar_t namlen, int hash)
931 {
932 	nc_hash_t *hp;
933 	ncache_t *ncp;
934 
935 	hp = &nc_hash[hash & nc_hashmask];
936 
937 	for (ncp = hp->hash_next; ncp != (ncache_t *)hp; ncp = ncp->hash_next) {
938 		if (ncp->hash == hash &&
939 		    ncp->dp == dp &&
940 		    ncp->namlen == namlen &&
941 		    bcmp(ncp->name, name, namlen) == 0)
942 			return (ncp);
943 	}
944 	return (NULL);
945 }
946 
947 #if ((1 << NBBY) - 1) < (MAXNAMELEN - 1)
948 #error ncache_t name length representation is too small
949 #endif
950 
951 void
dnlc_reduce_cache(void * reduce_percent)952 dnlc_reduce_cache(void *reduce_percent)
953 {
954 	if (dnlc_reduce_idle && (dnlc_nentries >= ncsize || reduce_percent)) {
955 		dnlc_reduce_idle = 0;
956 		if ((taskq_dispatch(system_taskq, do_dnlc_reduce_cache,
957 		    reduce_percent, TQ_NOSLEEP)) == TASKQID_INVALID)
958 			dnlc_reduce_idle = 1;
959 	}
960 }
961 
962 /*
963  * Get a new name cache entry.
964  * If the dnlc_reduce_cache() taskq isn't keeping up with demand, or memory
965  * is short then just return NULL. If we're over ncsize then kick off a
966  * thread to free some in use entries down to dnlc_nentries_low_water.
967  * Caller must initialise all fields except namlen.
968  * Component names are defined to be less than MAXNAMELEN
969  * which includes a null.
970  */
971 static ncache_t *
dnlc_get(uchar_t namlen)972 dnlc_get(uchar_t namlen)
973 {
974 	ncache_t *ncp;
975 
976 	if (dnlc_nentries > dnlc_max_nentries) {
977 		dnlc_max_nentries_cnt++; /* keep a statistic */
978 		return (NULL);
979 	}
980 	ncp = kmem_alloc(NCACHE_SIZE(namlen), KM_NOSLEEP);
981 	if (ncp == NULL) {
982 		return (NULL);
983 	}
984 	ncp->namlen = namlen;
985 	atomic_inc_32(&dnlc_nentries);
986 	dnlc_reduce_cache(NULL);
987 	return (ncp);
988 }
989 
990 /*
991  * Taskq routine to free up name cache entries to reduce the
992  * cache size to the low water mark if "reduce_percent" is not provided.
993  * If "reduce_percent" is provided, reduce cache size by
994  * (ncsize_onepercent * reduce_percent).
995  */
996 /*ARGSUSED*/
997 static void
do_dnlc_reduce_cache(void * reduce_percent)998 do_dnlc_reduce_cache(void *reduce_percent)
999 {
1000 	nc_hash_t *hp = dnlc_free_rotor, *start_hp = hp;
1001 	vnode_t *vp;
1002 	ncache_t *ncp;
1003 	int cnt;
1004 	uint_t low_water = dnlc_nentries_low_water;
1005 
1006 	if (reduce_percent) {
1007 		uint_t reduce_cnt;
1008 
1009 		/*
1010 		 * Never try to reduce the current number
1011 		 * of cache entries below 3% of ncsize.
1012 		 */
1013 		if (dnlc_nentries <= ncsize_min_percent) {
1014 			dnlc_reduce_idle = 1;
1015 			return;
1016 		}
1017 		reduce_cnt = ncsize_onepercent *
1018 		    (uint_t)(uintptr_t)reduce_percent;
1019 
1020 		if (reduce_cnt > dnlc_nentries ||
1021 		    dnlc_nentries - reduce_cnt < ncsize_min_percent)
1022 			low_water = ncsize_min_percent;
1023 		else
1024 			low_water = dnlc_nentries - reduce_cnt;
1025 	}
1026 
1027 	do {
1028 		/*
1029 		 * Find the first non empty hash queue without locking.
1030 		 * Only look at each hash queue once to avoid an infinite loop.
1031 		 */
1032 		do {
1033 			if (++hp == &nc_hash[nc_hashsz])
1034 				hp = nc_hash;
1035 		} while (hp->hash_next == (ncache_t *)hp && hp != start_hp);
1036 
1037 		/* return if all hash queues are empty. */
1038 		if (hp->hash_next == (ncache_t *)hp) {
1039 			dnlc_reduce_idle = 1;
1040 			return;
1041 		}
1042 
1043 		mutex_enter(&hp->hash_lock);
1044 		for (cnt = 0, ncp = hp->hash_prev; ncp != (ncache_t *)hp;
1045 		    ncp = ncp->hash_prev, cnt++) {
1046 			vp = ncp->vp;
1047 			/*
1048 			 * A name cache entry with a reference count
1049 			 * of one is only referenced by the dnlc.
1050 			 * Also negative cache entries are purged first.
1051 			 */
1052 			if (!vn_has_cached_data(vp) &&
1053 			    ((vp->v_count == 1) || (vp == DNLC_NO_VNODE))) {
1054 				ncs.ncs_pick_heur.value.ui64++;
1055 				goto found;
1056 			}
1057 			/*
1058 			 * Remove from the end of the chain if the
1059 			 * chain is too long
1060 			 */
1061 			if (cnt > dnlc_long_chain) {
1062 				ncp = hp->hash_prev;
1063 				ncs.ncs_pick_last.value.ui64++;
1064 				vp = ncp->vp;
1065 				goto found;
1066 			}
1067 		}
1068 		/* check for race and continue */
1069 		if (hp->hash_next == (ncache_t *)hp) {
1070 			mutex_exit(&hp->hash_lock);
1071 			continue;
1072 		}
1073 
1074 		ncp = hp->hash_prev; /* pick the last one in the hash queue */
1075 		ncs.ncs_pick_last.value.ui64++;
1076 		vp = ncp->vp;
1077 found:
1078 		/*
1079 		 * Remove from hash chain.
1080 		 */
1081 		nc_rmhash(ncp);
1082 		mutex_exit(&hp->hash_lock);
1083 		VN_RELE_DNLC(vp);
1084 		VN_RELE_DNLC(ncp->dp);
1085 		dnlc_free(ncp);
1086 	} while (dnlc_nentries > low_water);
1087 
1088 	dnlc_free_rotor = hp;
1089 	dnlc_reduce_idle = 1;
1090 }
1091 
1092 /*
1093  * Directory caching routines
1094  * ==========================
1095  *
1096  * See dnlc.h for details of the interfaces below.
1097  */
1098 
1099 /*
1100  * Lookup up an entry in a complete or partial directory cache.
1101  */
1102 dcret_t
dnlc_dir_lookup(dcanchor_t * dcap,const char * name,uint64_t * handle)1103 dnlc_dir_lookup(dcanchor_t *dcap, const char *name, uint64_t *handle)
1104 {
1105 	dircache_t *dcp;
1106 	dcentry_t *dep;
1107 	int hash;
1108 	int ret;
1109 	uchar_t namlen;
1110 
1111 	/*
1112 	 * can test without lock as we are only a cache
1113 	 */
1114 	if (!VALID_DIR_CACHE(dcap->dca_dircache)) {
1115 		ncs.ncs_dir_misses.value.ui64++;
1116 		return (DNOCACHE);
1117 	}
1118 
1119 	if (!dnlc_dir_enable) {
1120 		return (DNOCACHE);
1121 	}
1122 
1123 	mutex_enter(&dcap->dca_lock);
1124 	dcp = (dircache_t *)dcap->dca_dircache;
1125 	if (VALID_DIR_CACHE(dcp)) {
1126 		dcp->dc_actime = ddi_get_lbolt64();
1127 		DNLC_DIR_HASH(name, hash, namlen);
1128 		dep = dcp->dc_namehash[hash & dcp->dc_nhash_mask];
1129 		while (dep != NULL) {
1130 			if ((dep->de_hash == hash) &&
1131 			    (namlen == dep->de_namelen) &&
1132 			    bcmp(dep->de_name, name, namlen) == 0) {
1133 				*handle = dep->de_handle;
1134 				mutex_exit(&dcap->dca_lock);
1135 				ncs.ncs_dir_hits.value.ui64++;
1136 				return (DFOUND);
1137 			}
1138 			dep = dep->de_next;
1139 		}
1140 		if (dcp->dc_complete) {
1141 			ret = DNOENT;
1142 		} else {
1143 			ret = DNOCACHE;
1144 		}
1145 		mutex_exit(&dcap->dca_lock);
1146 		return (ret);
1147 	} else {
1148 		mutex_exit(&dcap->dca_lock);
1149 		ncs.ncs_dir_misses.value.ui64++;
1150 		return (DNOCACHE);
1151 	}
1152 }
1153 
1154 /*
1155  * Start a new directory cache. An estimate of the number of
1156  * entries is provided to as a quick check to ensure the directory
1157  * is cacheable.
1158  */
1159 dcret_t
dnlc_dir_start(dcanchor_t * dcap,uint_t num_entries)1160 dnlc_dir_start(dcanchor_t *dcap, uint_t num_entries)
1161 {
1162 	dircache_t *dcp;
1163 
1164 	if (!dnlc_dir_enable ||
1165 	    (num_entries < dnlc_dir_min_size)) {
1166 		return (DNOCACHE);
1167 	}
1168 
1169 	if (num_entries > dnlc_dir_max_size) {
1170 		return (DTOOBIG);
1171 	}
1172 
1173 	mutex_enter(&dc_head.dch_lock);
1174 	mutex_enter(&dcap->dca_lock);
1175 
1176 	if (dcap->dca_dircache == DC_RET_LOW_MEM) {
1177 		dcap->dca_dircache = NULL;
1178 		mutex_exit(&dcap->dca_lock);
1179 		mutex_exit(&dc_head.dch_lock);
1180 		return (DNOMEM);
1181 	}
1182 
1183 	/*
1184 	 * Check if there's currently a cache.
1185 	 * This probably only occurs on a race.
1186 	 */
1187 	if (dcap->dca_dircache != NULL) {
1188 		mutex_exit(&dcap->dca_lock);
1189 		mutex_exit(&dc_head.dch_lock);
1190 		return (DNOCACHE);
1191 	}
1192 
1193 	/*
1194 	 * Allocate the dircache struct, entry and free space hash tables.
1195 	 * These tables are initially just one entry but dynamically resize
1196 	 * when entries and free space are added or removed.
1197 	 */
1198 	if ((dcp = kmem_zalloc(sizeof (dircache_t), KM_NOSLEEP)) == NULL) {
1199 		goto error;
1200 	}
1201 	if ((dcp->dc_namehash = kmem_zalloc(sizeof (dcentry_t *),
1202 	    KM_NOSLEEP)) == NULL) {
1203 		goto error;
1204 	}
1205 	if ((dcp->dc_freehash = kmem_zalloc(sizeof (dcfree_t *),
1206 	    KM_NOSLEEP)) == NULL) {
1207 		goto error;
1208 	}
1209 
1210 	dcp->dc_anchor = dcap; /* set back pointer to anchor */
1211 	dcap->dca_dircache = dcp;
1212 
1213 	/* add into head of global chain */
1214 	dcp->dc_next = dc_head.dch_next;
1215 	dcp->dc_prev = (dircache_t *)&dc_head;
1216 	dcp->dc_next->dc_prev = dcp;
1217 	dc_head.dch_next = dcp;
1218 
1219 	mutex_exit(&dcap->dca_lock);
1220 	mutex_exit(&dc_head.dch_lock);
1221 	ncs.ncs_cur_dirs.value.ui64++;
1222 	ncs.ncs_dirs_cached.value.ui64++;
1223 	return (DOK);
1224 error:
1225 	if (dcp != NULL) {
1226 		if (dcp->dc_namehash) {
1227 			kmem_free(dcp->dc_namehash, sizeof (dcentry_t *));
1228 		}
1229 		kmem_free(dcp, sizeof (dircache_t));
1230 	}
1231 	/*
1232 	 * Must also kmem_free dcp->dc_freehash if more error cases are added
1233 	 */
1234 	mutex_exit(&dcap->dca_lock);
1235 	mutex_exit(&dc_head.dch_lock);
1236 	ncs.ncs_dir_start_nm.value.ui64++;
1237 	return (DNOCACHE);
1238 }
1239 
1240 /*
1241  * Add a directopry entry to a partial or complete directory cache.
1242  */
1243 dcret_t
dnlc_dir_add_entry(dcanchor_t * dcap,const char * name,uint64_t handle)1244 dnlc_dir_add_entry(dcanchor_t *dcap, const char *name, uint64_t handle)
1245 {
1246 	dircache_t *dcp;
1247 	dcentry_t **hp, *dep;
1248 	int hash;
1249 	uint_t capacity;
1250 	uchar_t namlen;
1251 
1252 	/*
1253 	 * Allocate the dcentry struct, including the variable
1254 	 * size name. Note, the null terminator is not copied.
1255 	 *
1256 	 * We do this outside the lock to avoid possible deadlock if
1257 	 * dnlc_dir_reclaim() is called as a result of memory shortage.
1258 	 */
1259 	DNLC_DIR_HASH(name, hash, namlen);
1260 	dep = kmem_alloc(DCENTTRY_SIZE(namlen), KM_NOSLEEP);
1261 	if (dep == NULL) {
1262 #ifdef DEBUG
1263 		/*
1264 		 * The kmem allocator generates random failures for
1265 		 * KM_NOSLEEP calls (see KMEM_RANDOM_ALLOCATION_FAILURE)
1266 		 * So try again before we blow away a perfectly good cache.
1267 		 * This is done not to cover an error but purely for
1268 		 * performance running a debug kernel.
1269 		 * This random error only occurs in debug mode.
1270 		 */
1271 		dep = kmem_alloc(DCENTTRY_SIZE(namlen), KM_NOSLEEP);
1272 		if (dep != NULL)
1273 			goto ok;
1274 #endif
1275 		ncs.ncs_dir_add_nm.value.ui64++;
1276 		/*
1277 		 * Free a directory cache. This may be the one we are
1278 		 * called with.
1279 		 */
1280 		dnlc_dir_reclaim(NULL);
1281 		dep = kmem_alloc(DCENTTRY_SIZE(namlen), KM_NOSLEEP);
1282 		if (dep == NULL) {
1283 			/*
1284 			 * still no memory, better delete this cache
1285 			 */
1286 			mutex_enter(&dcap->dca_lock);
1287 			dcp = (dircache_t *)dcap->dca_dircache;
1288 			if (VALID_DIR_CACHE(dcp)) {
1289 				dnlc_dir_abort(dcp);
1290 				dcap->dca_dircache = DC_RET_LOW_MEM;
1291 			}
1292 			mutex_exit(&dcap->dca_lock);
1293 			ncs.ncs_dir_addabort.value.ui64++;
1294 			return (DNOCACHE);
1295 		}
1296 		/*
1297 		 * fall through as if the 1st kmem_alloc had worked
1298 		 */
1299 	}
1300 #ifdef DEBUG
1301 ok:
1302 #endif
1303 	mutex_enter(&dcap->dca_lock);
1304 	dcp = (dircache_t *)dcap->dca_dircache;
1305 	if (VALID_DIR_CACHE(dcp)) {
1306 		/*
1307 		 * If the total number of entries goes above the max
1308 		 * then free this cache
1309 		 */
1310 		if ((dcp->dc_num_entries + dcp->dc_num_free) >
1311 		    dnlc_dir_max_size) {
1312 			mutex_exit(&dcap->dca_lock);
1313 			dnlc_dir_purge(dcap);
1314 			kmem_free(dep, DCENTTRY_SIZE(namlen));
1315 			ncs.ncs_dir_add_max.value.ui64++;
1316 			return (DTOOBIG);
1317 		}
1318 		dcp->dc_num_entries++;
1319 		capacity = (dcp->dc_nhash_mask + 1) << dnlc_dir_hash_size_shift;
1320 		if (dcp->dc_num_entries >=
1321 		    (capacity << dnlc_dir_hash_resize_shift)) {
1322 			dnlc_dir_adjust_nhash(dcp);
1323 		}
1324 		hp = &dcp->dc_namehash[hash & dcp->dc_nhash_mask];
1325 
1326 		/*
1327 		 * Initialise and chain in new entry
1328 		 */
1329 		dep->de_handle = handle;
1330 		dep->de_hash = hash;
1331 		/*
1332 		 * Note de_namelen is a uchar_t to conserve space
1333 		 * and alignment padding. The max length of any
1334 		 * pathname component is defined as MAXNAMELEN
1335 		 * which is 256 (including the terminating null).
1336 		 * So provided this doesn't change, we don't include the null,
1337 		 * we always use bcmp to compare strings, and we don't
1338 		 * start storing full names, then we are ok.
1339 		 * The space savings is worth it.
1340 		 */
1341 		dep->de_namelen = namlen;
1342 		bcopy(name, dep->de_name, namlen);
1343 		dep->de_next = *hp;
1344 		*hp = dep;
1345 		dcp->dc_actime = ddi_get_lbolt64();
1346 		mutex_exit(&dcap->dca_lock);
1347 		ncs.ncs_dir_num_ents.value.ui64++;
1348 		return (DOK);
1349 	} else {
1350 		mutex_exit(&dcap->dca_lock);
1351 		kmem_free(dep, DCENTTRY_SIZE(namlen));
1352 		return (DNOCACHE);
1353 	}
1354 }
1355 
1356 /*
1357  * Add free space to a partial or complete directory cache.
1358  */
1359 dcret_t
dnlc_dir_add_space(dcanchor_t * dcap,uint_t len,uint64_t handle)1360 dnlc_dir_add_space(dcanchor_t *dcap, uint_t len, uint64_t handle)
1361 {
1362 	dircache_t *dcp;
1363 	dcfree_t *dfp, **hp;
1364 	uint_t capacity;
1365 
1366 	/*
1367 	 * We kmem_alloc outside the lock to avoid possible deadlock if
1368 	 * dnlc_dir_reclaim() is called as a result of memory shortage.
1369 	 */
1370 	dfp = kmem_cache_alloc(dnlc_dir_space_cache, KM_NOSLEEP);
1371 	if (dfp == NULL) {
1372 #ifdef DEBUG
1373 		/*
1374 		 * The kmem allocator generates random failures for
1375 		 * KM_NOSLEEP calls (see KMEM_RANDOM_ALLOCATION_FAILURE)
1376 		 * So try again before we blow away a perfectly good cache.
1377 		 * This random error only occurs in debug mode
1378 		 */
1379 		dfp = kmem_cache_alloc(dnlc_dir_space_cache, KM_NOSLEEP);
1380 		if (dfp != NULL)
1381 			goto ok;
1382 #endif
1383 		ncs.ncs_dir_add_nm.value.ui64++;
1384 		/*
1385 		 * Free a directory cache. This may be the one we are
1386 		 * called with.
1387 		 */
1388 		dnlc_dir_reclaim(NULL);
1389 		dfp = kmem_cache_alloc(dnlc_dir_space_cache, KM_NOSLEEP);
1390 		if (dfp == NULL) {
1391 			/*
1392 			 * still no memory, better delete this cache
1393 			 */
1394 			mutex_enter(&dcap->dca_lock);
1395 			dcp = (dircache_t *)dcap->dca_dircache;
1396 			if (VALID_DIR_CACHE(dcp)) {
1397 				dnlc_dir_abort(dcp);
1398 				dcap->dca_dircache = DC_RET_LOW_MEM;
1399 			}
1400 			mutex_exit(&dcap->dca_lock);
1401 			ncs.ncs_dir_addabort.value.ui64++;
1402 			return (DNOCACHE);
1403 		}
1404 		/*
1405 		 * fall through as if the 1st kmem_alloc had worked
1406 		 */
1407 	}
1408 
1409 #ifdef DEBUG
1410 ok:
1411 #endif
1412 	mutex_enter(&dcap->dca_lock);
1413 	dcp = (dircache_t *)dcap->dca_dircache;
1414 	if (VALID_DIR_CACHE(dcp)) {
1415 		if ((dcp->dc_num_entries + dcp->dc_num_free) >
1416 		    dnlc_dir_max_size) {
1417 			mutex_exit(&dcap->dca_lock);
1418 			dnlc_dir_purge(dcap);
1419 			kmem_cache_free(dnlc_dir_space_cache, dfp);
1420 			ncs.ncs_dir_add_max.value.ui64++;
1421 			return (DTOOBIG);
1422 		}
1423 		dcp->dc_num_free++;
1424 		capacity = (dcp->dc_fhash_mask + 1) << dnlc_dir_hash_size_shift;
1425 		if (dcp->dc_num_free >=
1426 		    (capacity << dnlc_dir_hash_resize_shift)) {
1427 			dnlc_dir_adjust_fhash(dcp);
1428 		}
1429 		/*
1430 		 * Initialise and chain a new entry
1431 		 */
1432 		dfp->df_handle = handle;
1433 		dfp->df_len = len;
1434 		dcp->dc_actime = ddi_get_lbolt64();
1435 		hp = &(dcp->dc_freehash[DDFHASH(handle, dcp)]);
1436 		dfp->df_next = *hp;
1437 		*hp = dfp;
1438 		mutex_exit(&dcap->dca_lock);
1439 		ncs.ncs_dir_num_ents.value.ui64++;
1440 		return (DOK);
1441 	} else {
1442 		mutex_exit(&dcap->dca_lock);
1443 		kmem_cache_free(dnlc_dir_space_cache, dfp);
1444 		return (DNOCACHE);
1445 	}
1446 }
1447 
1448 /*
1449  * Mark a directory cache as complete.
1450  */
1451 void
dnlc_dir_complete(dcanchor_t * dcap)1452 dnlc_dir_complete(dcanchor_t *dcap)
1453 {
1454 	dircache_t *dcp;
1455 
1456 	mutex_enter(&dcap->dca_lock);
1457 	dcp = (dircache_t *)dcap->dca_dircache;
1458 	if (VALID_DIR_CACHE(dcp)) {
1459 		dcp->dc_complete = B_TRUE;
1460 	}
1461 	mutex_exit(&dcap->dca_lock);
1462 }
1463 
1464 /*
1465  * Internal routine to delete a partial or full directory cache.
1466  * No additional locking needed.
1467  */
1468 static void
dnlc_dir_abort(dircache_t * dcp)1469 dnlc_dir_abort(dircache_t *dcp)
1470 {
1471 	dcentry_t *dep, *nhp;
1472 	dcfree_t *fep, *fhp;
1473 	uint_t nhtsize = dcp->dc_nhash_mask + 1; /* name hash table size */
1474 	uint_t fhtsize = dcp->dc_fhash_mask + 1; /* free hash table size */
1475 	uint_t i;
1476 
1477 	/*
1478 	 * Free up the cached name entries and hash table
1479 	 */
1480 	for (i = 0; i < nhtsize; i++) { /* for each hash bucket */
1481 		nhp = dcp->dc_namehash[i];
1482 		while (nhp != NULL) { /* for each chained entry */
1483 			dep = nhp->de_next;
1484 			kmem_free(nhp, DCENTTRY_SIZE(nhp->de_namelen));
1485 			nhp = dep;
1486 		}
1487 	}
1488 	kmem_free(dcp->dc_namehash, sizeof (dcentry_t *) * nhtsize);
1489 
1490 	/*
1491 	 * Free up the free space entries and hash table
1492 	 */
1493 	for (i = 0; i < fhtsize; i++) { /* for each hash bucket */
1494 		fhp = dcp->dc_freehash[i];
1495 		while (fhp != NULL) { /* for each chained entry */
1496 			fep = fhp->df_next;
1497 			kmem_cache_free(dnlc_dir_space_cache, fhp);
1498 			fhp = fep;
1499 		}
1500 	}
1501 	kmem_free(dcp->dc_freehash, sizeof (dcfree_t *) * fhtsize);
1502 
1503 	/*
1504 	 * Finally free the directory cache structure itself
1505 	 */
1506 	ncs.ncs_dir_num_ents.value.ui64 -= (dcp->dc_num_entries +
1507 	    dcp->dc_num_free);
1508 	kmem_free(dcp, sizeof (dircache_t));
1509 	ncs.ncs_cur_dirs.value.ui64--;
1510 }
1511 
1512 /*
1513  * Remove a partial or complete directory cache
1514  */
1515 void
dnlc_dir_purge(dcanchor_t * dcap)1516 dnlc_dir_purge(dcanchor_t *dcap)
1517 {
1518 	dircache_t *dcp;
1519 
1520 	mutex_enter(&dc_head.dch_lock);
1521 	mutex_enter(&dcap->dca_lock);
1522 	dcp = (dircache_t *)dcap->dca_dircache;
1523 	if (!VALID_DIR_CACHE(dcp)) {
1524 		mutex_exit(&dcap->dca_lock);
1525 		mutex_exit(&dc_head.dch_lock);
1526 		return;
1527 	}
1528 	dcap->dca_dircache = NULL;
1529 	/*
1530 	 * Unchain from global list
1531 	 */
1532 	dcp->dc_prev->dc_next = dcp->dc_next;
1533 	dcp->dc_next->dc_prev = dcp->dc_prev;
1534 	mutex_exit(&dcap->dca_lock);
1535 	mutex_exit(&dc_head.dch_lock);
1536 	dnlc_dir_abort(dcp);
1537 }
1538 
1539 /*
1540  * Remove an entry from a complete or partial directory cache.
1541  * Return the handle if it's non null.
1542  */
1543 dcret_t
dnlc_dir_rem_entry(dcanchor_t * dcap,const char * name,uint64_t * handlep)1544 dnlc_dir_rem_entry(dcanchor_t *dcap, const char *name, uint64_t *handlep)
1545 {
1546 	dircache_t *dcp;
1547 	dcentry_t **prevpp, *te;
1548 	uint_t capacity;
1549 	int hash;
1550 	int ret;
1551 	uchar_t namlen;
1552 
1553 	if (!dnlc_dir_enable) {
1554 		return (DNOCACHE);
1555 	}
1556 
1557 	mutex_enter(&dcap->dca_lock);
1558 	dcp = (dircache_t *)dcap->dca_dircache;
1559 	if (VALID_DIR_CACHE(dcp)) {
1560 		dcp->dc_actime = ddi_get_lbolt64();
1561 		if (dcp->dc_nhash_mask > 0) { /* ie not minimum */
1562 			capacity = (dcp->dc_nhash_mask + 1) <<
1563 			    dnlc_dir_hash_size_shift;
1564 			if (dcp->dc_num_entries <=
1565 			    (capacity >> dnlc_dir_hash_resize_shift)) {
1566 				dnlc_dir_adjust_nhash(dcp);
1567 			}
1568 		}
1569 		DNLC_DIR_HASH(name, hash, namlen);
1570 		prevpp = &dcp->dc_namehash[hash & dcp->dc_nhash_mask];
1571 		while (*prevpp != NULL) {
1572 			if (((*prevpp)->de_hash == hash) &&
1573 			    (namlen == (*prevpp)->de_namelen) &&
1574 			    bcmp((*prevpp)->de_name, name, namlen) == 0) {
1575 				if (handlep != NULL) {
1576 					*handlep = (*prevpp)->de_handle;
1577 				}
1578 				te = *prevpp;
1579 				*prevpp = (*prevpp)->de_next;
1580 				kmem_free(te, DCENTTRY_SIZE(te->de_namelen));
1581 
1582 				/*
1583 				 * If the total number of entries
1584 				 * falls below half the minimum number
1585 				 * of entries then free this cache.
1586 				 */
1587 				if (--dcp->dc_num_entries <
1588 				    (dnlc_dir_min_size >> 1)) {
1589 					mutex_exit(&dcap->dca_lock);
1590 					dnlc_dir_purge(dcap);
1591 				} else {
1592 					mutex_exit(&dcap->dca_lock);
1593 				}
1594 				ncs.ncs_dir_num_ents.value.ui64--;
1595 				return (DFOUND);
1596 			}
1597 			prevpp = &((*prevpp)->de_next);
1598 		}
1599 		if (dcp->dc_complete) {
1600 			ncs.ncs_dir_reme_fai.value.ui64++;
1601 			ret = DNOENT;
1602 		} else {
1603 			ret = DNOCACHE;
1604 		}
1605 		mutex_exit(&dcap->dca_lock);
1606 		return (ret);
1607 	} else {
1608 		mutex_exit(&dcap->dca_lock);
1609 		return (DNOCACHE);
1610 	}
1611 }
1612 
1613 
1614 /*
1615  * Remove free space of at least the given length from a complete
1616  * or partial directory cache.
1617  */
1618 dcret_t
dnlc_dir_rem_space_by_len(dcanchor_t * dcap,uint_t len,uint64_t * handlep)1619 dnlc_dir_rem_space_by_len(dcanchor_t *dcap, uint_t len, uint64_t *handlep)
1620 {
1621 	dircache_t *dcp;
1622 	dcfree_t **prevpp, *tfp;
1623 	uint_t fhtsize; /* free hash table size */
1624 	uint_t i;
1625 	uint_t capacity;
1626 	int ret;
1627 
1628 	if (!dnlc_dir_enable) {
1629 		return (DNOCACHE);
1630 	}
1631 
1632 	mutex_enter(&dcap->dca_lock);
1633 	dcp = (dircache_t *)dcap->dca_dircache;
1634 	if (VALID_DIR_CACHE(dcp)) {
1635 		dcp->dc_actime = ddi_get_lbolt64();
1636 		if (dcp->dc_fhash_mask > 0) { /* ie not minimum */
1637 			capacity = (dcp->dc_fhash_mask + 1) <<
1638 			    dnlc_dir_hash_size_shift;
1639 			if (dcp->dc_num_free <=
1640 			    (capacity >> dnlc_dir_hash_resize_shift)) {
1641 				dnlc_dir_adjust_fhash(dcp);
1642 			}
1643 		}
1644 		/*
1645 		 * Search for an entry of the appropriate size
1646 		 * on a first fit basis.
1647 		 */
1648 		fhtsize = dcp->dc_fhash_mask + 1;
1649 		for (i = 0; i < fhtsize; i++) { /* for each hash bucket */
1650 			prevpp = &(dcp->dc_freehash[i]);
1651 			while (*prevpp != NULL) {
1652 				if ((*prevpp)->df_len >= len) {
1653 					*handlep = (*prevpp)->df_handle;
1654 					tfp = *prevpp;
1655 					*prevpp = (*prevpp)->df_next;
1656 					dcp->dc_num_free--;
1657 					mutex_exit(&dcap->dca_lock);
1658 					kmem_cache_free(dnlc_dir_space_cache,
1659 					    tfp);
1660 					ncs.ncs_dir_num_ents.value.ui64--;
1661 					return (DFOUND);
1662 				}
1663 				prevpp = &((*prevpp)->df_next);
1664 			}
1665 		}
1666 		if (dcp->dc_complete) {
1667 			ret = DNOENT;
1668 		} else {
1669 			ret = DNOCACHE;
1670 		}
1671 		mutex_exit(&dcap->dca_lock);
1672 		return (ret);
1673 	} else {
1674 		mutex_exit(&dcap->dca_lock);
1675 		return (DNOCACHE);
1676 	}
1677 }
1678 
1679 /*
1680  * Remove free space with the given handle from a complete or partial
1681  * directory cache.
1682  */
1683 dcret_t
dnlc_dir_rem_space_by_handle(dcanchor_t * dcap,uint64_t handle)1684 dnlc_dir_rem_space_by_handle(dcanchor_t *dcap, uint64_t handle)
1685 {
1686 	dircache_t *dcp;
1687 	dcfree_t **prevpp, *tfp;
1688 	uint_t capacity;
1689 	int ret;
1690 
1691 	if (!dnlc_dir_enable) {
1692 		return (DNOCACHE);
1693 	}
1694 
1695 	mutex_enter(&dcap->dca_lock);
1696 	dcp = (dircache_t *)dcap->dca_dircache;
1697 	if (VALID_DIR_CACHE(dcp)) {
1698 		dcp->dc_actime = ddi_get_lbolt64();
1699 		if (dcp->dc_fhash_mask > 0) { /* ie not minimum */
1700 			capacity = (dcp->dc_fhash_mask + 1) <<
1701 			    dnlc_dir_hash_size_shift;
1702 			if (dcp->dc_num_free <=
1703 			    (capacity >> dnlc_dir_hash_resize_shift)) {
1704 				dnlc_dir_adjust_fhash(dcp);
1705 			}
1706 		}
1707 
1708 		/*
1709 		 * search for the exact entry
1710 		 */
1711 		prevpp = &(dcp->dc_freehash[DDFHASH(handle, dcp)]);
1712 		while (*prevpp != NULL) {
1713 			if ((*prevpp)->df_handle == handle) {
1714 				tfp = *prevpp;
1715 				*prevpp = (*prevpp)->df_next;
1716 				dcp->dc_num_free--;
1717 				mutex_exit(&dcap->dca_lock);
1718 				kmem_cache_free(dnlc_dir_space_cache, tfp);
1719 				ncs.ncs_dir_num_ents.value.ui64--;
1720 				return (DFOUND);
1721 			}
1722 			prevpp = &((*prevpp)->df_next);
1723 		}
1724 		if (dcp->dc_complete) {
1725 			ncs.ncs_dir_rems_fai.value.ui64++;
1726 			ret = DNOENT;
1727 		} else {
1728 			ret = DNOCACHE;
1729 		}
1730 		mutex_exit(&dcap->dca_lock);
1731 		return (ret);
1732 	} else {
1733 		mutex_exit(&dcap->dca_lock);
1734 		return (DNOCACHE);
1735 	}
1736 }
1737 
1738 /*
1739  * Update the handle of an directory cache entry.
1740  */
1741 dcret_t
dnlc_dir_update(dcanchor_t * dcap,const char * name,uint64_t handle)1742 dnlc_dir_update(dcanchor_t *dcap, const char *name, uint64_t handle)
1743 {
1744 	dircache_t *dcp;
1745 	dcentry_t *dep;
1746 	int hash;
1747 	int ret;
1748 	uchar_t namlen;
1749 
1750 	if (!dnlc_dir_enable) {
1751 		return (DNOCACHE);
1752 	}
1753 
1754 	mutex_enter(&dcap->dca_lock);
1755 	dcp = (dircache_t *)dcap->dca_dircache;
1756 	if (VALID_DIR_CACHE(dcp)) {
1757 		dcp->dc_actime = ddi_get_lbolt64();
1758 		DNLC_DIR_HASH(name, hash, namlen);
1759 		dep = dcp->dc_namehash[hash & dcp->dc_nhash_mask];
1760 		while (dep != NULL) {
1761 			if ((dep->de_hash == hash) &&
1762 			    (namlen == dep->de_namelen) &&
1763 			    bcmp(dep->de_name, name, namlen) == 0) {
1764 				dep->de_handle = handle;
1765 				mutex_exit(&dcap->dca_lock);
1766 				return (DFOUND);
1767 			}
1768 			dep = dep->de_next;
1769 		}
1770 		if (dcp->dc_complete) {
1771 			ncs.ncs_dir_upd_fail.value.ui64++;
1772 			ret = DNOENT;
1773 		} else {
1774 			ret = DNOCACHE;
1775 		}
1776 		mutex_exit(&dcap->dca_lock);
1777 		return (ret);
1778 	} else {
1779 		mutex_exit(&dcap->dca_lock);
1780 		return (DNOCACHE);
1781 	}
1782 }
1783 
1784 void
dnlc_dir_fini(dcanchor_t * dcap)1785 dnlc_dir_fini(dcanchor_t *dcap)
1786 {
1787 	dircache_t *dcp;
1788 
1789 	mutex_enter(&dc_head.dch_lock);
1790 	mutex_enter(&dcap->dca_lock);
1791 	dcp = (dircache_t *)dcap->dca_dircache;
1792 	if (VALID_DIR_CACHE(dcp)) {
1793 		/*
1794 		 * Unchain from global list
1795 		 */
1796 		ncs.ncs_dir_finipurg.value.ui64++;
1797 		dcp->dc_prev->dc_next = dcp->dc_next;
1798 		dcp->dc_next->dc_prev = dcp->dc_prev;
1799 	} else {
1800 		dcp = NULL;
1801 	}
1802 	dcap->dca_dircache = NULL;
1803 	mutex_exit(&dcap->dca_lock);
1804 	mutex_exit(&dc_head.dch_lock);
1805 	mutex_destroy(&dcap->dca_lock);
1806 	if (dcp) {
1807 		dnlc_dir_abort(dcp);
1808 	}
1809 }
1810 
1811 /*
1812  * Reclaim callback for dnlc directory caching.
1813  * Invoked by the kernel memory allocator when memory gets tight.
1814  * This is a pretty serious condition and can lead easily lead to system
1815  * hangs if not enough space is returned.
1816  *
1817  * Deciding which directory (or directories) to purge is tricky.
1818  * Purging everything is an overkill, but purging just the oldest used
1819  * was found to lead to hangs. The largest cached directories use the
1820  * most memory, but take the most effort to rebuild, whereas the smaller
1821  * ones have little value and give back little space. So what to do?
1822  *
1823  * The current policy is to continue purging the oldest used directories
1824  * until at least dnlc_dir_min_reclaim directory entries have been purged.
1825  */
1826 /*ARGSUSED*/
1827 static void
dnlc_dir_reclaim(void * unused)1828 dnlc_dir_reclaim(void *unused)
1829 {
1830 	dircache_t *dcp, *oldest;
1831 	uint_t dirent_cnt = 0;
1832 
1833 	mutex_enter(&dc_head.dch_lock);
1834 	while (dirent_cnt < dnlc_dir_min_reclaim) {
1835 		dcp = dc_head.dch_next;
1836 		oldest = NULL;
1837 		while (dcp != (dircache_t *)&dc_head) {
1838 			if (oldest == NULL) {
1839 				oldest = dcp;
1840 			} else {
1841 				if (dcp->dc_actime < oldest->dc_actime) {
1842 					oldest = dcp;
1843 				}
1844 			}
1845 			dcp = dcp->dc_next;
1846 		}
1847 		if (oldest == NULL) {
1848 			/* nothing to delete */
1849 			mutex_exit(&dc_head.dch_lock);
1850 			return;
1851 		}
1852 		/*
1853 		 * remove from directory chain and purge
1854 		 */
1855 		oldest->dc_prev->dc_next = oldest->dc_next;
1856 		oldest->dc_next->dc_prev = oldest->dc_prev;
1857 		mutex_enter(&oldest->dc_anchor->dca_lock);
1858 		/*
1859 		 * If this was the last entry then it must be too large.
1860 		 * Mark it as such by saving a special dircache_t
1861 		 * pointer (DC_RET_LOW_MEM) in the anchor. The error DNOMEM
1862 		 * will be presented to the caller of dnlc_dir_start()
1863 		 */
1864 		if (oldest->dc_next == oldest->dc_prev) {
1865 			oldest->dc_anchor->dca_dircache = DC_RET_LOW_MEM;
1866 			ncs.ncs_dir_rec_last.value.ui64++;
1867 		} else {
1868 			oldest->dc_anchor->dca_dircache = NULL;
1869 			ncs.ncs_dir_recl_any.value.ui64++;
1870 		}
1871 		mutex_exit(&oldest->dc_anchor->dca_lock);
1872 		dirent_cnt += oldest->dc_num_entries;
1873 		dnlc_dir_abort(oldest);
1874 	}
1875 	mutex_exit(&dc_head.dch_lock);
1876 }
1877 
1878 /*
1879  * Dynamically grow or shrink the size of the name hash table
1880  */
1881 static void
dnlc_dir_adjust_nhash(dircache_t * dcp)1882 dnlc_dir_adjust_nhash(dircache_t *dcp)
1883 {
1884 	dcentry_t **newhash, *dep, **nhp, *tep;
1885 	uint_t newsize;
1886 	uint_t oldsize;
1887 	uint_t newsizemask;
1888 	int i;
1889 
1890 	/*
1891 	 * Allocate new hash table
1892 	 */
1893 	newsize = dcp->dc_num_entries >> dnlc_dir_hash_size_shift;
1894 	newhash = kmem_zalloc(sizeof (dcentry_t *) * newsize, KM_NOSLEEP);
1895 	if (newhash == NULL) {
1896 		/*
1897 		 * System is short on memory just return
1898 		 * Note, the old hash table is still usable.
1899 		 * This return is unlikely to repeatedy occur, because
1900 		 * either some other directory caches will be reclaimed
1901 		 * due to memory shortage, thus freeing memory, or this
1902 		 * directory cahe will be reclaimed.
1903 		 */
1904 		return;
1905 	}
1906 	oldsize = dcp->dc_nhash_mask + 1;
1907 	dcp->dc_nhash_mask = newsizemask = newsize - 1;
1908 
1909 	/*
1910 	 * Move entries from the old table to the new
1911 	 */
1912 	for (i = 0; i < oldsize; i++) { /* for each hash bucket */
1913 		dep = dcp->dc_namehash[i];
1914 		while (dep != NULL) { /* for each chained entry */
1915 			tep = dep;
1916 			dep = dep->de_next;
1917 			nhp = &newhash[tep->de_hash & newsizemask];
1918 			tep->de_next = *nhp;
1919 			*nhp = tep;
1920 		}
1921 	}
1922 
1923 	/*
1924 	 * delete old hash table and set new one in place
1925 	 */
1926 	kmem_free(dcp->dc_namehash, sizeof (dcentry_t *) * oldsize);
1927 	dcp->dc_namehash = newhash;
1928 }
1929 
1930 /*
1931  * Dynamically grow or shrink the size of the free space hash table
1932  */
1933 static void
dnlc_dir_adjust_fhash(dircache_t * dcp)1934 dnlc_dir_adjust_fhash(dircache_t *dcp)
1935 {
1936 	dcfree_t **newhash, *dfp, **nhp, *tfp;
1937 	uint_t newsize;
1938 	uint_t oldsize;
1939 	int i;
1940 
1941 	/*
1942 	 * Allocate new hash table
1943 	 */
1944 	newsize = dcp->dc_num_free >> dnlc_dir_hash_size_shift;
1945 	newhash = kmem_zalloc(sizeof (dcfree_t *) * newsize, KM_NOSLEEP);
1946 	if (newhash == NULL) {
1947 		/*
1948 		 * System is short on memory just return
1949 		 * Note, the old hash table is still usable.
1950 		 * This return is unlikely to repeatedy occur, because
1951 		 * either some other directory caches will be reclaimed
1952 		 * due to memory shortage, thus freeing memory, or this
1953 		 * directory cahe will be reclaimed.
1954 		 */
1955 		return;
1956 	}
1957 	oldsize = dcp->dc_fhash_mask + 1;
1958 	dcp->dc_fhash_mask = newsize - 1;
1959 
1960 	/*
1961 	 * Move entries from the old table to the new
1962 	 */
1963 	for (i = 0; i < oldsize; i++) { /* for each hash bucket */
1964 		dfp = dcp->dc_freehash[i];
1965 		while (dfp != NULL) { /* for each chained entry */
1966 			tfp = dfp;
1967 			dfp = dfp->df_next;
1968 			nhp = &newhash[DDFHASH(tfp->df_handle, dcp)];
1969 			tfp->df_next = *nhp;
1970 			*nhp = tfp;
1971 		}
1972 	}
1973 
1974 	/*
1975 	 * delete old hash table and set new one in place
1976 	 */
1977 	kmem_free(dcp->dc_freehash, sizeof (dcfree_t *) * oldsize);
1978 	dcp->dc_freehash = newhash;
1979 }
1980