xref: /titanic_41/usr/src/uts/common/fs/dnlc.c (revision 0aea656903145d947d28737ae11caddf89707ec8)
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  */
25 
26 /*	Copyright (c) 1983, 1984, 1985, 1986, 1987, 1988, 1989 AT&T	*/
27 /*	  All Rights Reserved  	*/
28 
29 /*
30  * University Copyright- Copyright (c) 1982, 1986, 1988
31  * The Regents of the University of California
32  * All Rights Reserved
33  *
34  * University Acknowledgment- Portions of this document are derived from
35  * software developed by the University of California, Berkeley, and its
36  * contributors.
37  */
38 
39 #include <sys/types.h>
40 #include <sys/systm.h>
41 #include <sys/param.h>
42 #include <sys/t_lock.h>
43 #include <sys/systm.h>
44 #include <sys/vfs.h>
45 #include <sys/vnode.h>
46 #include <sys/dnlc.h>
47 #include <sys/kmem.h>
48 #include <sys/cmn_err.h>
49 #include <sys/vtrace.h>
50 #include <sys/bitmap.h>
51 #include <sys/var.h>
52 #include <sys/sysmacros.h>
53 #include <sys/kstat.h>
54 #include <sys/atomic.h>
55 #include <sys/taskq.h>
56 
57 /*
58  * Directory name lookup cache.
59  * Based on code originally done by Robert Elz at Melbourne.
60  *
61  * Names found by directory scans are retained in a cache
62  * for future reference.  Each hash chain is ordered by LRU
63  * Cache is indexed by hash value obtained from (vp, name)
64  * where the vp refers to the directory containing the name.
65  */
66 
67 /*
68  * We want to be able to identify files that are referenced only by the DNLC.
69  * When adding a reference from the DNLC, call VN_HOLD_DNLC instead of VN_HOLD,
70  * since multiple DNLC references should only be counted once in v_count. The
71  * VN_HOLD macro itself is aliased to VN_HOLD_CALLER in this file to help
72  * differentiate the behaviors.  (Unfortunately it is not possible to #undef
73  * VN_HOLD and retain VN_HOLD_CALLER. Ideally a Makefile rule would grep
74  * uncommented C tokens to check that VN_HOLD is referenced only once in this
75  * file, to define VN_HOLD_CALLER.)
76  */
77 #define	VN_HOLD_CALLER	VN_HOLD
78 #define	VN_HOLD_DNLC(vp)	{	\
79 	mutex_enter(&(vp)->v_lock);	\
80 	if ((vp)->v_count_dnlc == 0)	\
81 		(vp)->v_count++;	\
82 	(vp)->v_count_dnlc++;		\
83 	mutex_exit(&(vp)->v_lock);	\
84 }
85 #define	VN_RELE_DNLC(vp)	{	\
86 	vn_rele_dnlc(vp);		\
87 }
88 
89 /*
90  * Tunable nc_hashavelen is the average length desired for this chain, from
91  * which the size of the nc_hash table is derived at create time.
92  */
93 #define	NC_HASHAVELEN_DEFAULT	4
94 int nc_hashavelen = NC_HASHAVELEN_DEFAULT;
95 
96 /*
97  * NC_MOVETOFRONT is the move-to-front threshold: if the hash lookup
98  * depth exceeds this value, we move the looked-up entry to the front of
99  * its hash chain.  The idea is to make sure that the most frequently
100  * accessed entries are found most quickly (by keeping them near the
101  * front of their hash chains).
102  */
103 #define	NC_MOVETOFRONT	2
104 
105 /*
106  *
107  * DNLC_MAX_RELE is used to size an array on the stack when releasing
108  * vnodes. This array is used rather than calling VN_RELE() inline because
109  * all dnlc locks must be dropped by that time in order to avoid a
110  * possible deadlock. This deadlock occurs when the dnlc holds the last
111  * reference to the vnode and so the VOP_INACTIVE vector is called which
112  * can in turn call back into the dnlc. A global array was used but had
113  * many problems:
114  *	1) Actually doesn't have an upper bound on the array size as
115  *	   entries can be added after starting the purge.
116  *	2) The locking scheme causes a hang.
117  *	3) Caused serialisation on the global lock.
118  *	4) The array was often unnecessarily huge.
119  *
120  * Note the current value 8 allows up to 4 cache entries (to be purged
121  * from each hash chain), before having to cycle around and retry.
122  * This ought to be ample given that nc_hashavelen is typically very small.
123  */
124 #define	DNLC_MAX_RELE	8 /* must be even */
125 
126 /*
127  * Hash table of name cache entries for fast lookup, dynamically
128  * allocated at startup.
129  */
130 nc_hash_t *nc_hash;
131 
132 /*
133  * Rotors. Used to select entries on a round-robin basis.
134  */
135 static nc_hash_t *dnlc_purge_fs1_rotor;
136 static nc_hash_t *dnlc_free_rotor;
137 
138 /*
139  * # of dnlc entries (uninitialized)
140  *
141  * the initial value was chosen as being
142  * a random string of bits, probably not
143  * normally chosen by a systems administrator
144  */
145 int ncsize = -1;
146 volatile uint32_t dnlc_nentries = 0;	/* current num of name cache entries */
147 static int nc_hashsz;			/* size of hash table */
148 static int nc_hashmask;			/* size of hash table minus 1 */
149 
150 /*
151  * The dnlc_reduce_cache() taskq queue is activated when there are
152  * ncsize name cache entries and if no parameter is provided, it reduces
153  * the size down to dnlc_nentries_low_water, which is by default one
154  * hundreth less (or 99%) of ncsize.
155  *
156  * If a parameter is provided to dnlc_reduce_cache(), then we reduce
157  * the size down based on ncsize_onepercent - where ncsize_onepercent
158  * is 1% of ncsize; however, we never let dnlc_reduce_cache() reduce
159  * the size below 3% of ncsize (ncsize_min_percent).
160  */
161 #define	DNLC_LOW_WATER_DIVISOR_DEFAULT 100
162 uint_t dnlc_low_water_divisor = DNLC_LOW_WATER_DIVISOR_DEFAULT;
163 uint_t dnlc_nentries_low_water;
164 int dnlc_reduce_idle = 1; /* no locking needed */
165 uint_t ncsize_onepercent;
166 uint_t ncsize_min_percent;
167 
168 /*
169  * If dnlc_nentries hits dnlc_max_nentries (twice ncsize)
170  * then this means the dnlc_reduce_cache() taskq is failing to
171  * keep up. In this case we refuse to add new entries to the dnlc
172  * until the taskq catches up.
173  */
174 uint_t dnlc_max_nentries; /* twice ncsize */
175 uint64_t dnlc_max_nentries_cnt = 0; /* statistic on times we failed */
176 
177 /*
178  * Tunable to define when we should just remove items from
179  * the end of the chain.
180  */
181 #define	DNLC_LONG_CHAIN 8
182 uint_t dnlc_long_chain = DNLC_LONG_CHAIN;
183 
184 /*
185  * ncstats has been deprecated, due to the integer size of the counters
186  * which can easily overflow in the dnlc.
187  * It is maintained (at some expense) for compatability.
188  * The preferred interface is the kstat accessible nc_stats below.
189  */
190 struct ncstats ncstats;
191 
192 struct nc_stats ncs = {
193 	{ "hits",			KSTAT_DATA_UINT64 },
194 	{ "misses",			KSTAT_DATA_UINT64 },
195 	{ "negative_cache_hits",	KSTAT_DATA_UINT64 },
196 	{ "enters",			KSTAT_DATA_UINT64 },
197 	{ "double_enters",		KSTAT_DATA_UINT64 },
198 	{ "purge_total_entries",	KSTAT_DATA_UINT64 },
199 	{ "purge_all",			KSTAT_DATA_UINT64 },
200 	{ "purge_vp",			KSTAT_DATA_UINT64 },
201 	{ "purge_vfs",			KSTAT_DATA_UINT64 },
202 	{ "purge_fs1",			KSTAT_DATA_UINT64 },
203 	{ "pick_free",			KSTAT_DATA_UINT64 },
204 	{ "pick_heuristic",		KSTAT_DATA_UINT64 },
205 	{ "pick_last",			KSTAT_DATA_UINT64 },
206 
207 	/* directory caching stats */
208 
209 	{ "dir_hits",			KSTAT_DATA_UINT64 },
210 	{ "dir_misses",			KSTAT_DATA_UINT64 },
211 	{ "dir_cached_current",		KSTAT_DATA_UINT64 },
212 	{ "dir_entries_cached_current",	KSTAT_DATA_UINT64 },
213 	{ "dir_cached_total",		KSTAT_DATA_UINT64 },
214 	{ "dir_start_no_memory",	KSTAT_DATA_UINT64 },
215 	{ "dir_add_no_memory",		KSTAT_DATA_UINT64 },
216 	{ "dir_add_abort",		KSTAT_DATA_UINT64 },
217 	{ "dir_add_max",		KSTAT_DATA_UINT64 },
218 	{ "dir_remove_entry_fail",	KSTAT_DATA_UINT64 },
219 	{ "dir_remove_space_fail",	KSTAT_DATA_UINT64 },
220 	{ "dir_update_fail",		KSTAT_DATA_UINT64 },
221 	{ "dir_fini_purge",		KSTAT_DATA_UINT64 },
222 	{ "dir_reclaim_last",		KSTAT_DATA_UINT64 },
223 	{ "dir_reclaim_any",		KSTAT_DATA_UINT64 },
224 };
225 
226 static int doingcache = 1;
227 
228 vnode_t negative_cache_vnode;
229 
230 /*
231  * Insert entry at the front of the queue
232  */
233 #define	nc_inshash(ncp, hp) \
234 { \
235 	(ncp)->hash_next = (hp)->hash_next; \
236 	(ncp)->hash_prev = (ncache_t *)(hp); \
237 	(hp)->hash_next->hash_prev = (ncp); \
238 	(hp)->hash_next = (ncp); \
239 }
240 
241 /*
242  * Remove entry from hash queue
243  */
244 #define	nc_rmhash(ncp) \
245 { \
246 	(ncp)->hash_prev->hash_next = (ncp)->hash_next; \
247 	(ncp)->hash_next->hash_prev = (ncp)->hash_prev; \
248 	(ncp)->hash_prev = NULL; \
249 	(ncp)->hash_next = NULL; \
250 }
251 
252 /*
253  * Free an entry.
254  */
255 #define	dnlc_free(ncp) \
256 { \
257 	kmem_free((ncp), sizeof (ncache_t) + (ncp)->namlen); \
258 	atomic_dec_32(&dnlc_nentries); \
259 }
260 
261 
262 /*
263  * Cached directory info.
264  * ======================
265  */
266 
267 /*
268  * Cached directory free space hash function.
269  * Needs the free space handle and the dcp to get the hash table size
270  * Returns the hash index.
271  */
272 #define	DDFHASH(handle, dcp) ((handle >> 2) & (dcp)->dc_fhash_mask)
273 
274 /*
275  * Cached directory name entry hash function.
276  * Uses the name and returns in the input arguments the hash and the name
277  * length.
278  */
279 #define	DNLC_DIR_HASH(name, hash, namelen)			\
280 	{							\
281 		char Xc;					\
282 		const char *Xcp;				\
283 		hash = *name;					\
284 		for (Xcp = (name + 1); (Xc = *Xcp) != 0; Xcp++)	\
285 			hash = (hash << 4) + hash + Xc;		\
286 		ASSERT((Xcp - (name)) <= ((1 << NBBY) - 1));	\
287 		namelen = Xcp - (name);				\
288 	}
289 
290 /* special dircache_t pointer to indicate error should be returned */
291 /*
292  * The anchor directory cache pointer can contain 3 types of values,
293  * 1) NULL: No directory cache
294  * 2) DC_RET_LOW_MEM (-1): There was a directory cache that found to be
295  *    too big or a memory shortage occurred. This value remains in the
296  *    pointer until a dnlc_dir_start() which returns the a DNOMEM error.
297  *    This is kludgy but efficient and only visible in this source file.
298  * 3) A valid cache pointer.
299  */
300 #define	DC_RET_LOW_MEM (dircache_t *)1
301 #define	VALID_DIR_CACHE(dcp) ((dircache_t *)(dcp) > DC_RET_LOW_MEM)
302 
303 /* Tunables */
304 uint_t dnlc_dir_enable = 1; /* disable caching directories by setting to 0 */
305 uint_t dnlc_dir_min_size = 40; /* min no of directory entries before caching */
306 uint_t dnlc_dir_max_size = UINT_MAX; /* ditto maximum */
307 uint_t dnlc_dir_hash_size_shift = 3; /* 8 entries per hash bucket */
308 uint_t dnlc_dir_min_reclaim =  350000; /* approx 1MB of dcentrys */
309 /*
310  * dnlc_dir_hash_resize_shift determines when the hash tables
311  * get re-adjusted due to growth or shrinkage
312  * - currently 2 indicating that there can be at most 4
313  * times or at least one quarter the number of entries
314  * before hash table readjustment. Note that with
315  * dnlc_dir_hash_size_shift above set at 3 this would
316  * mean readjustment would occur if the average number
317  * of entries went above 32 or below 2
318  */
319 uint_t dnlc_dir_hash_resize_shift = 2; /* readjust rate */
320 
321 static kmem_cache_t *dnlc_dir_space_cache; /* free space entry cache */
322 static dchead_t dc_head; /* anchor of cached directories */
323 
324 /* Prototypes */
325 static ncache_t *dnlc_get(uchar_t namlen);
326 static ncache_t *dnlc_search(vnode_t *dp, const char *name, uchar_t namlen,
327     int hash);
328 static void dnlc_dir_reclaim(void *unused);
329 static void dnlc_dir_abort(dircache_t *dcp);
330 static void dnlc_dir_adjust_fhash(dircache_t *dcp);
331 static void dnlc_dir_adjust_nhash(dircache_t *dcp);
332 static void do_dnlc_reduce_cache(void *);
333 
334 
335 /*
336  * Initialize the directory cache.
337  */
338 void
dnlc_init()339 dnlc_init()
340 {
341 	nc_hash_t *hp;
342 	kstat_t *ksp;
343 	int i;
344 
345 	/*
346 	 * Set up the size of the dnlc (ncsize) and its low water mark.
347 	 */
348 	if (ncsize == -1) {
349 		/* calculate a reasonable size for the low water */
350 		dnlc_nentries_low_water = 4 * (v.v_proc + maxusers) + 320;
351 		ncsize = dnlc_nentries_low_water +
352 		    (dnlc_nentries_low_water / dnlc_low_water_divisor);
353 	} else {
354 		/* don't change the user specified ncsize */
355 		dnlc_nentries_low_water =
356 		    ncsize - (ncsize / dnlc_low_water_divisor);
357 	}
358 	if (ncsize <= 0) {
359 		doingcache = 0;
360 		dnlc_dir_enable = 0; /* also disable directory caching */
361 		ncsize = 0;
362 		cmn_err(CE_NOTE, "name cache (dnlc) disabled");
363 		return;
364 	}
365 	dnlc_max_nentries = ncsize * 2;
366 	ncsize_onepercent = ncsize / 100;
367 	ncsize_min_percent = ncsize_onepercent * 3;
368 
369 	/*
370 	 * Initialise the hash table.
371 	 * Compute hash size rounding to the next power of two.
372 	 */
373 	nc_hashsz = ncsize / nc_hashavelen;
374 	nc_hashsz = 1 << highbit(nc_hashsz);
375 	nc_hashmask = nc_hashsz - 1;
376 	nc_hash = kmem_zalloc(nc_hashsz * sizeof (*nc_hash), KM_SLEEP);
377 	for (i = 0; i < nc_hashsz; i++) {
378 		hp = (nc_hash_t *)&nc_hash[i];
379 		mutex_init(&hp->hash_lock, NULL, MUTEX_DEFAULT, NULL);
380 		hp->hash_next = (ncache_t *)hp;
381 		hp->hash_prev = (ncache_t *)hp;
382 	}
383 
384 	/*
385 	 * Initialize rotors
386 	 */
387 	dnlc_free_rotor = dnlc_purge_fs1_rotor = &nc_hash[0];
388 
389 	/*
390 	 * Set up the directory caching to use kmem_cache_alloc
391 	 * for its free space entries so that we can get a callback
392 	 * when the system is short on memory, to allow us to free
393 	 * up some memory. we don't use the constructor/deconstructor
394 	 * functions.
395 	 */
396 	dnlc_dir_space_cache = kmem_cache_create("dnlc_space_cache",
397 	    sizeof (dcfree_t), 0, NULL, NULL, dnlc_dir_reclaim, NULL,
398 	    NULL, 0);
399 
400 	/*
401 	 * Initialise the head of the cached directory structures
402 	 */
403 	mutex_init(&dc_head.dch_lock, NULL, MUTEX_DEFAULT, NULL);
404 	dc_head.dch_next = (dircache_t *)&dc_head;
405 	dc_head.dch_prev = (dircache_t *)&dc_head;
406 
407 	/*
408 	 * Initialise the reference count of the negative cache vnode to 1
409 	 * so that it never goes away (VOP_INACTIVE isn't called on it).
410 	 */
411 	negative_cache_vnode.v_count = 1;
412 	negative_cache_vnode.v_count_dnlc = 0;
413 
414 	/*
415 	 * Initialise kstats - both the old compatability raw kind and
416 	 * the more extensive named stats.
417 	 */
418 	ksp = kstat_create("unix", 0, "ncstats", "misc", KSTAT_TYPE_RAW,
419 	    sizeof (struct ncstats), KSTAT_FLAG_VIRTUAL);
420 	if (ksp) {
421 		ksp->ks_data = (void *) &ncstats;
422 		kstat_install(ksp);
423 	}
424 	ksp = kstat_create("unix", 0, "dnlcstats", "misc", KSTAT_TYPE_NAMED,
425 	    sizeof (ncs) / sizeof (kstat_named_t), KSTAT_FLAG_VIRTUAL);
426 	if (ksp) {
427 		ksp->ks_data = (void *) &ncs;
428 		kstat_install(ksp);
429 	}
430 }
431 
432 /*
433  * Add a name to the directory cache.
434  */
435 void
dnlc_enter(vnode_t * dp,const char * name,vnode_t * vp)436 dnlc_enter(vnode_t *dp, const char *name, vnode_t *vp)
437 {
438 	ncache_t *ncp;
439 	nc_hash_t *hp;
440 	uchar_t namlen;
441 	int hash;
442 
443 	TRACE_0(TR_FAC_NFS, TR_DNLC_ENTER_START, "dnlc_enter_start:");
444 
445 	if (!doingcache) {
446 		TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
447 		    "dnlc_enter_end:(%S) %d", "not caching", 0);
448 		return;
449 	}
450 
451 	/*
452 	 * Get a new dnlc entry. Assume the entry won't be in the cache
453 	 * and initialize it now
454 	 */
455 	DNLCHASH(name, dp, hash, namlen);
456 	if ((ncp = dnlc_get(namlen)) == NULL)
457 		return;
458 	ncp->dp = dp;
459 	VN_HOLD_DNLC(dp);
460 	ncp->vp = vp;
461 	VN_HOLD_DNLC(vp);
462 	bcopy(name, ncp->name, namlen + 1); /* name and null */
463 	ncp->hash = hash;
464 	hp = &nc_hash[hash & nc_hashmask];
465 
466 	mutex_enter(&hp->hash_lock);
467 	if (dnlc_search(dp, name, namlen, hash) != NULL) {
468 		mutex_exit(&hp->hash_lock);
469 		ncstats.dbl_enters++;
470 		ncs.ncs_dbl_enters.value.ui64++;
471 		VN_RELE_DNLC(dp);
472 		VN_RELE_DNLC(vp);
473 		dnlc_free(ncp);		/* crfree done here */
474 		TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
475 		    "dnlc_enter_end:(%S) %d", "dbl enter", ncstats.dbl_enters);
476 		return;
477 	}
478 	/*
479 	 * Insert back into the hash chain.
480 	 */
481 	nc_inshash(ncp, hp);
482 	mutex_exit(&hp->hash_lock);
483 	ncstats.enters++;
484 	ncs.ncs_enters.value.ui64++;
485 	TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
486 	    "dnlc_enter_end:(%S) %d", "done", ncstats.enters);
487 }
488 
489 /*
490  * Add a name to the directory cache.
491  *
492  * This function is basically identical with
493  * dnlc_enter().  The difference is that when the
494  * desired dnlc entry is found, the vnode in the
495  * ncache is compared with the vnode passed in.
496  *
497  * If they are not equal then the ncache is
498  * updated with the passed in vnode.  Otherwise
499  * it just frees up the newly allocated dnlc entry.
500  */
501 void
dnlc_update(vnode_t * dp,const char * name,vnode_t * vp)502 dnlc_update(vnode_t *dp, const char *name, vnode_t *vp)
503 {
504 	ncache_t *ncp;
505 	ncache_t *tcp;
506 	vnode_t *tvp;
507 	nc_hash_t *hp;
508 	int hash;
509 	uchar_t namlen;
510 
511 	TRACE_0(TR_FAC_NFS, TR_DNLC_ENTER_START, "dnlc_update_start:");
512 
513 	if (!doingcache) {
514 		TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
515 		    "dnlc_update_end:(%S) %d", "not caching", 0);
516 		return;
517 	}
518 
519 	/*
520 	 * Get a new dnlc entry and initialize it now.
521 	 * If we fail to get a new entry, call dnlc_remove() to purge
522 	 * any existing dnlc entry including negative cache (DNLC_NO_VNODE)
523 	 * entry.
524 	 * Failure to clear an existing entry could result in false dnlc
525 	 * lookup (negative/stale entry).
526 	 */
527 	DNLCHASH(name, dp, hash, namlen);
528 	if ((ncp = dnlc_get(namlen)) == NULL) {
529 		dnlc_remove(dp, name);
530 		return;
531 	}
532 	ncp->dp = dp;
533 	VN_HOLD_DNLC(dp);
534 	ncp->vp = vp;
535 	VN_HOLD_DNLC(vp);
536 	bcopy(name, ncp->name, namlen + 1); /* name and null */
537 	ncp->hash = hash;
538 	hp = &nc_hash[hash & nc_hashmask];
539 
540 	mutex_enter(&hp->hash_lock);
541 	if ((tcp = dnlc_search(dp, name, namlen, hash)) != NULL) {
542 		if (tcp->vp != vp) {
543 			tvp = tcp->vp;
544 			tcp->vp = vp;
545 			mutex_exit(&hp->hash_lock);
546 			VN_RELE_DNLC(tvp);
547 			ncstats.enters++;
548 			ncs.ncs_enters.value.ui64++;
549 			TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
550 			    "dnlc_update_end:(%S) %d", "done", ncstats.enters);
551 		} else {
552 			mutex_exit(&hp->hash_lock);
553 			VN_RELE_DNLC(vp);
554 			ncstats.dbl_enters++;
555 			ncs.ncs_dbl_enters.value.ui64++;
556 			TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
557 			    "dnlc_update_end:(%S) %d",
558 			    "dbl enter", ncstats.dbl_enters);
559 		}
560 		VN_RELE_DNLC(dp);
561 		dnlc_free(ncp);		/* crfree done here */
562 		return;
563 	}
564 	/*
565 	 * insert the new entry, since it is not in dnlc yet
566 	 */
567 	nc_inshash(ncp, hp);
568 	mutex_exit(&hp->hash_lock);
569 	ncstats.enters++;
570 	ncs.ncs_enters.value.ui64++;
571 	TRACE_2(TR_FAC_NFS, TR_DNLC_ENTER_END,
572 	    "dnlc_update_end:(%S) %d", "done", ncstats.enters);
573 }
574 
575 /*
576  * Look up a name in the directory name cache.
577  *
578  * Return a doubly-held vnode if found: one hold so that it may
579  * remain in the cache for other users, the other hold so that
580  * the cache is not re-cycled and the identity of the vnode is
581  * lost before the caller can use the vnode.
582  */
583 vnode_t *
dnlc_lookup(vnode_t * dp,const char * name)584 dnlc_lookup(vnode_t *dp, const char *name)
585 {
586 	ncache_t *ncp;
587 	nc_hash_t *hp;
588 	vnode_t *vp;
589 	int hash, depth;
590 	uchar_t namlen;
591 
592 	TRACE_2(TR_FAC_NFS, TR_DNLC_LOOKUP_START,
593 	    "dnlc_lookup_start:dp %x name %s", dp, name);
594 
595 	if (!doingcache) {
596 		TRACE_4(TR_FAC_NFS, TR_DNLC_LOOKUP_END,
597 		    "dnlc_lookup_end:%S %d vp %x name %s",
598 		    "not_caching", 0, NULL, name);
599 		return (NULL);
600 	}
601 
602 	DNLCHASH(name, dp, hash, namlen);
603 	depth = 1;
604 	hp = &nc_hash[hash & nc_hashmask];
605 	mutex_enter(&hp->hash_lock);
606 
607 	for (ncp = hp->hash_next; ncp != (ncache_t *)hp;
608 	    ncp = ncp->hash_next) {
609 		if (ncp->hash == hash &&	/* fast signature check */
610 		    ncp->dp == dp &&
611 		    ncp->namlen == namlen &&
612 		    bcmp(ncp->name, name, namlen) == 0) {
613 			/*
614 			 * Move this entry to the head of its hash chain
615 			 * if it's not already close.
616 			 */
617 			if (depth > NC_MOVETOFRONT) {
618 				ncache_t *next = ncp->hash_next;
619 				ncache_t *prev = ncp->hash_prev;
620 
621 				prev->hash_next = next;
622 				next->hash_prev = prev;
623 				ncp->hash_next = next = hp->hash_next;
624 				ncp->hash_prev = (ncache_t *)hp;
625 				next->hash_prev = ncp;
626 				hp->hash_next = ncp;
627 
628 				ncstats.move_to_front++;
629 			}
630 
631 			/*
632 			 * Put a hold on the vnode now so its identity
633 			 * can't change before the caller has a chance to
634 			 * put a hold on it.
635 			 */
636 			vp = ncp->vp;
637 			VN_HOLD_CALLER(vp);
638 			mutex_exit(&hp->hash_lock);
639 			ncstats.hits++;
640 			ncs.ncs_hits.value.ui64++;
641 			if (vp == DNLC_NO_VNODE) {
642 				ncs.ncs_neg_hits.value.ui64++;
643 			}
644 			TRACE_4(TR_FAC_NFS, TR_DNLC_LOOKUP_END,
645 			    "dnlc_lookup_end:%S %d vp %x name %s", "hit",
646 			    ncstats.hits, vp, name);
647 			return (vp);
648 		}
649 		depth++;
650 	}
651 
652 	mutex_exit(&hp->hash_lock);
653 	ncstats.misses++;
654 	ncs.ncs_misses.value.ui64++;
655 	TRACE_4(TR_FAC_NFS, TR_DNLC_LOOKUP_END,
656 	    "dnlc_lookup_end:%S %d vp %x name %s", "miss", ncstats.misses,
657 	    NULL, name);
658 	return (NULL);
659 }
660 
661 /*
662  * Remove an entry in the directory name cache.
663  */
664 void
dnlc_remove(vnode_t * dp,const char * name)665 dnlc_remove(vnode_t *dp, const char *name)
666 {
667 	ncache_t *ncp;
668 	nc_hash_t *hp;
669 	uchar_t namlen;
670 	int hash;
671 
672 	if (!doingcache)
673 		return;
674 	DNLCHASH(name, dp, hash, namlen);
675 	hp = &nc_hash[hash & nc_hashmask];
676 
677 	mutex_enter(&hp->hash_lock);
678 	if (ncp = dnlc_search(dp, name, namlen, hash)) {
679 		/*
680 		 * Free up the entry
681 		 */
682 		nc_rmhash(ncp);
683 		mutex_exit(&hp->hash_lock);
684 		VN_RELE_DNLC(ncp->vp);
685 		VN_RELE_DNLC(ncp->dp);
686 		dnlc_free(ncp);
687 		return;
688 	}
689 	mutex_exit(&hp->hash_lock);
690 }
691 
692 /*
693  * Purge the entire cache.
694  */
695 void
dnlc_purge()696 dnlc_purge()
697 {
698 	nc_hash_t *nch;
699 	ncache_t *ncp;
700 	int index;
701 	int i;
702 	vnode_t *nc_rele[DNLC_MAX_RELE];
703 
704 	if (!doingcache)
705 		return;
706 
707 	ncstats.purges++;
708 	ncs.ncs_purge_all.value.ui64++;
709 
710 	for (nch = nc_hash; nch < &nc_hash[nc_hashsz]; nch++) {
711 		index = 0;
712 		mutex_enter(&nch->hash_lock);
713 		ncp = nch->hash_next;
714 		while (ncp != (ncache_t *)nch) {
715 			ncache_t *np;
716 
717 			np = ncp->hash_next;
718 			nc_rele[index++] = ncp->vp;
719 			nc_rele[index++] = ncp->dp;
720 
721 			nc_rmhash(ncp);
722 			dnlc_free(ncp);
723 			ncp = np;
724 			ncs.ncs_purge_total.value.ui64++;
725 			if (index == DNLC_MAX_RELE)
726 				break;
727 		}
728 		mutex_exit(&nch->hash_lock);
729 
730 		/* Release holds on all the vnodes now that we have no locks */
731 		for (i = 0; i < index; i++) {
732 			VN_RELE_DNLC(nc_rele[i]);
733 		}
734 		if (ncp != (ncache_t *)nch) {
735 			nch--; /* Do current hash chain again */
736 		}
737 	}
738 }
739 
740 /*
741  * Purge any cache entries referencing a vnode. Exit as soon as the dnlc
742  * reference count goes to zero (the caller still holds a reference).
743  */
744 void
dnlc_purge_vp(vnode_t * vp)745 dnlc_purge_vp(vnode_t *vp)
746 {
747 	nc_hash_t *nch;
748 	ncache_t *ncp;
749 	int index;
750 	vnode_t *nc_rele[DNLC_MAX_RELE];
751 
752 	ASSERT(vp->v_count > 0);
753 	if (vp->v_count_dnlc == 0) {
754 		return;
755 	}
756 
757 	if (!doingcache)
758 		return;
759 
760 	ncstats.purges++;
761 	ncs.ncs_purge_vp.value.ui64++;
762 
763 	for (nch = nc_hash; nch < &nc_hash[nc_hashsz]; nch++) {
764 		index = 0;
765 		mutex_enter(&nch->hash_lock);
766 		ncp = nch->hash_next;
767 		while (ncp != (ncache_t *)nch) {
768 			ncache_t *np;
769 
770 			np = ncp->hash_next;
771 			if (ncp->dp == vp || ncp->vp == vp) {
772 				nc_rele[index++] = ncp->vp;
773 				nc_rele[index++] = ncp->dp;
774 				nc_rmhash(ncp);
775 				dnlc_free(ncp);
776 				ncs.ncs_purge_total.value.ui64++;
777 				if (index == DNLC_MAX_RELE) {
778 					ncp = np;
779 					break;
780 				}
781 			}
782 			ncp = np;
783 		}
784 		mutex_exit(&nch->hash_lock);
785 
786 		/* Release holds on all the vnodes now that we have no locks */
787 		while (index) {
788 			VN_RELE_DNLC(nc_rele[--index]);
789 		}
790 
791 		if (vp->v_count_dnlc == 0) {
792 			return;
793 		}
794 
795 		if (ncp != (ncache_t *)nch) {
796 			nch--; /* Do current hash chain again */
797 		}
798 	}
799 }
800 
801 /*
802  * Purge cache entries referencing a vfsp.  Caller supplies a count
803  * of entries to purge; up to that many will be freed.  A count of
804  * zero indicates that all such entries should be purged.  Returns
805  * the number of entries that were purged.
806  */
807 int
dnlc_purge_vfsp(vfs_t * vfsp,int count)808 dnlc_purge_vfsp(vfs_t *vfsp, int count)
809 {
810 	nc_hash_t *nch;
811 	ncache_t *ncp;
812 	int n = 0;
813 	int index;
814 	int i;
815 	vnode_t *nc_rele[DNLC_MAX_RELE];
816 
817 	if (!doingcache)
818 		return (0);
819 
820 	ncstats.purges++;
821 	ncs.ncs_purge_vfs.value.ui64++;
822 
823 	for (nch = nc_hash; nch < &nc_hash[nc_hashsz]; nch++) {
824 		index = 0;
825 		mutex_enter(&nch->hash_lock);
826 		ncp = nch->hash_next;
827 		while (ncp != (ncache_t *)nch) {
828 			ncache_t *np;
829 
830 			np = ncp->hash_next;
831 			ASSERT(ncp->dp != NULL);
832 			ASSERT(ncp->vp != NULL);
833 			if ((ncp->dp->v_vfsp == vfsp) ||
834 			    (ncp->vp->v_vfsp == vfsp)) {
835 				n++;
836 				nc_rele[index++] = ncp->vp;
837 				nc_rele[index++] = ncp->dp;
838 				nc_rmhash(ncp);
839 				dnlc_free(ncp);
840 				ncs.ncs_purge_total.value.ui64++;
841 				if (index == DNLC_MAX_RELE) {
842 					ncp = np;
843 					break;
844 				}
845 				if (count != 0 && n >= count) {
846 					break;
847 				}
848 			}
849 			ncp = np;
850 		}
851 		mutex_exit(&nch->hash_lock);
852 		/* Release holds on all the vnodes now that we have no locks */
853 		for (i = 0; i < index; i++) {
854 			VN_RELE_DNLC(nc_rele[i]);
855 		}
856 		if (count != 0 && n >= count) {
857 			return (n);
858 		}
859 		if (ncp != (ncache_t *)nch) {
860 			nch--; /* Do current hash chain again */
861 		}
862 	}
863 	return (n);
864 }
865 
866 /*
867  * Purge 1 entry from the dnlc that is part of the filesystem(s)
868  * represented by 'vop'. The purpose of this routine is to allow
869  * users of the dnlc to free a vnode that is being held by the dnlc.
870  *
871  * If we find a vnode that we release which will result in
872  * freeing the underlying vnode (count was 1), return 1, 0
873  * if no appropriate vnodes found.
874  *
875  * Note, vop is not the 'right' identifier for a filesystem.
876  */
877 int
dnlc_fs_purge1(vnodeops_t * vop)878 dnlc_fs_purge1(vnodeops_t *vop)
879 {
880 	nc_hash_t *end;
881 	nc_hash_t *hp;
882 	ncache_t *ncp;
883 	vnode_t *vp;
884 
885 	if (!doingcache)
886 		return (0);
887 
888 	ncs.ncs_purge_fs1.value.ui64++;
889 
890 	/*
891 	 * Scan the dnlc entries looking for a likely candidate.
892 	 */
893 	hp = end = dnlc_purge_fs1_rotor;
894 
895 	do {
896 		if (++hp == &nc_hash[nc_hashsz])
897 			hp = nc_hash;
898 		dnlc_purge_fs1_rotor = hp;
899 		if (hp->hash_next == (ncache_t *)hp)
900 			continue;
901 		mutex_enter(&hp->hash_lock);
902 		for (ncp = hp->hash_prev;
903 		    ncp != (ncache_t *)hp;
904 		    ncp = ncp->hash_prev) {
905 			vp = ncp->vp;
906 			if (!vn_has_cached_data(vp) && (vp->v_count == 1) &&
907 			    vn_matchops(vp, vop))
908 				break;
909 		}
910 		if (ncp != (ncache_t *)hp) {
911 			nc_rmhash(ncp);
912 			mutex_exit(&hp->hash_lock);
913 			VN_RELE_DNLC(ncp->dp);
914 			VN_RELE_DNLC(vp)
915 			dnlc_free(ncp);
916 			ncs.ncs_purge_total.value.ui64++;
917 			return (1);
918 		}
919 		mutex_exit(&hp->hash_lock);
920 	} while (hp != end);
921 	return (0);
922 }
923 
924 /*
925  * Utility routine to search for a cache entry. Return the
926  * ncache entry if found, NULL otherwise.
927  */
928 static ncache_t *
dnlc_search(vnode_t * dp,const char * name,uchar_t namlen,int hash)929 dnlc_search(vnode_t *dp, const char *name, uchar_t namlen, int hash)
930 {
931 	nc_hash_t *hp;
932 	ncache_t *ncp;
933 
934 	hp = &nc_hash[hash & nc_hashmask];
935 
936 	for (ncp = hp->hash_next; ncp != (ncache_t *)hp; ncp = ncp->hash_next) {
937 		if (ncp->hash == hash &&
938 		    ncp->dp == dp &&
939 		    ncp->namlen == namlen &&
940 		    bcmp(ncp->name, name, namlen) == 0)
941 			return (ncp);
942 	}
943 	return (NULL);
944 }
945 
946 #if ((1 << NBBY) - 1) < (MAXNAMELEN - 1)
947 #error ncache_t name length representation is too small
948 #endif
949 
950 void
dnlc_reduce_cache(void * reduce_percent)951 dnlc_reduce_cache(void *reduce_percent)
952 {
953 	if (dnlc_reduce_idle && (dnlc_nentries >= ncsize || reduce_percent)) {
954 		dnlc_reduce_idle = 0;
955 		if ((taskq_dispatch(system_taskq, do_dnlc_reduce_cache,
956 		    reduce_percent, TQ_NOSLEEP)) == NULL)
957 			dnlc_reduce_idle = 1;
958 	}
959 }
960 
961 /*
962  * Get a new name cache entry.
963  * If the dnlc_reduce_cache() taskq isn't keeping up with demand, or memory
964  * is short then just return NULL. If we're over ncsize then kick off a
965  * thread to free some in use entries down to dnlc_nentries_low_water.
966  * Caller must initialise all fields except namlen.
967  * Component names are defined to be less than MAXNAMELEN
968  * which includes a null.
969  */
970 static ncache_t *
dnlc_get(uchar_t namlen)971 dnlc_get(uchar_t namlen)
972 {
973 	ncache_t *ncp;
974 
975 	if (dnlc_nentries > dnlc_max_nentries) {
976 		dnlc_max_nentries_cnt++; /* keep a statistic */
977 		return (NULL);
978 	}
979 	ncp = kmem_alloc(sizeof (ncache_t) + namlen, KM_NOSLEEP);
980 	if (ncp == NULL) {
981 		return (NULL);
982 	}
983 	ncp->namlen = namlen;
984 	atomic_inc_32(&dnlc_nentries);
985 	dnlc_reduce_cache(NULL);
986 	return (ncp);
987 }
988 
989 /*
990  * Taskq routine to free up name cache entries to reduce the
991  * cache size to the low water mark if "reduce_percent" is not provided.
992  * If "reduce_percent" is provided, reduce cache size by
993  * (ncsize_onepercent * reduce_percent).
994  */
995 /*ARGSUSED*/
996 static void
do_dnlc_reduce_cache(void * reduce_percent)997 do_dnlc_reduce_cache(void *reduce_percent)
998 {
999 	nc_hash_t *hp = dnlc_free_rotor, *start_hp = hp;
1000 	vnode_t *vp;
1001 	ncache_t *ncp;
1002 	int cnt;
1003 	uint_t low_water = dnlc_nentries_low_water;
1004 
1005 	if (reduce_percent) {
1006 		uint_t reduce_cnt;
1007 
1008 		/*
1009 		 * Never try to reduce the current number
1010 		 * of cache entries below 3% of ncsize.
1011 		 */
1012 		if (dnlc_nentries <= ncsize_min_percent) {
1013 			dnlc_reduce_idle = 1;
1014 			return;
1015 		}
1016 		reduce_cnt = ncsize_onepercent *
1017 		    (uint_t)(uintptr_t)reduce_percent;
1018 
1019 		if (reduce_cnt > dnlc_nentries ||
1020 		    dnlc_nentries - reduce_cnt < ncsize_min_percent)
1021 			low_water = ncsize_min_percent;
1022 		else
1023 			low_water = dnlc_nentries - reduce_cnt;
1024 	}
1025 
1026 	do {
1027 		/*
1028 		 * Find the first non empty hash queue without locking.
1029 		 * Only look at each hash queue once to avoid an infinite loop.
1030 		 */
1031 		do {
1032 			if (++hp == &nc_hash[nc_hashsz])
1033 				hp = nc_hash;
1034 		} while (hp->hash_next == (ncache_t *)hp && hp != start_hp);
1035 
1036 		/* return if all hash queues are empty. */
1037 		if (hp->hash_next == (ncache_t *)hp) {
1038 			dnlc_reduce_idle = 1;
1039 			return;
1040 		}
1041 
1042 		mutex_enter(&hp->hash_lock);
1043 		for (cnt = 0, ncp = hp->hash_prev; ncp != (ncache_t *)hp;
1044 		    ncp = ncp->hash_prev, cnt++) {
1045 			vp = ncp->vp;
1046 			/*
1047 			 * A name cache entry with a reference count
1048 			 * of one is only referenced by the dnlc.
1049 			 * Also negative cache entries are purged first.
1050 			 */
1051 			if (!vn_has_cached_data(vp) &&
1052 			    ((vp->v_count == 1) || (vp == DNLC_NO_VNODE))) {
1053 				ncs.ncs_pick_heur.value.ui64++;
1054 				goto found;
1055 			}
1056 			/*
1057 			 * Remove from the end of the chain if the
1058 			 * chain is too long
1059 			 */
1060 			if (cnt > dnlc_long_chain) {
1061 				ncp = hp->hash_prev;
1062 				ncs.ncs_pick_last.value.ui64++;
1063 				vp = ncp->vp;
1064 				goto found;
1065 			}
1066 		}
1067 		/* check for race and continue */
1068 		if (hp->hash_next == (ncache_t *)hp) {
1069 			mutex_exit(&hp->hash_lock);
1070 			continue;
1071 		}
1072 
1073 		ncp = hp->hash_prev; /* pick the last one in the hash queue */
1074 		ncs.ncs_pick_last.value.ui64++;
1075 		vp = ncp->vp;
1076 found:
1077 		/*
1078 		 * Remove from hash chain.
1079 		 */
1080 		nc_rmhash(ncp);
1081 		mutex_exit(&hp->hash_lock);
1082 		VN_RELE_DNLC(vp);
1083 		VN_RELE_DNLC(ncp->dp);
1084 		dnlc_free(ncp);
1085 	} while (dnlc_nentries > low_water);
1086 
1087 	dnlc_free_rotor = hp;
1088 	dnlc_reduce_idle = 1;
1089 }
1090 
1091 /*
1092  * Directory caching routines
1093  * ==========================
1094  *
1095  * See dnlc.h for details of the interfaces below.
1096  */
1097 
1098 /*
1099  * Lookup up an entry in a complete or partial directory cache.
1100  */
1101 dcret_t
dnlc_dir_lookup(dcanchor_t * dcap,const char * name,uint64_t * handle)1102 dnlc_dir_lookup(dcanchor_t *dcap, const char *name, uint64_t *handle)
1103 {
1104 	dircache_t *dcp;
1105 	dcentry_t *dep;
1106 	int hash;
1107 	int ret;
1108 	uchar_t namlen;
1109 
1110 	/*
1111 	 * can test without lock as we are only a cache
1112 	 */
1113 	if (!VALID_DIR_CACHE(dcap->dca_dircache)) {
1114 		ncs.ncs_dir_misses.value.ui64++;
1115 		return (DNOCACHE);
1116 	}
1117 
1118 	if (!dnlc_dir_enable) {
1119 		return (DNOCACHE);
1120 	}
1121 
1122 	mutex_enter(&dcap->dca_lock);
1123 	dcp = (dircache_t *)dcap->dca_dircache;
1124 	if (VALID_DIR_CACHE(dcp)) {
1125 		dcp->dc_actime = ddi_get_lbolt64();
1126 		DNLC_DIR_HASH(name, hash, namlen);
1127 		dep = dcp->dc_namehash[hash & dcp->dc_nhash_mask];
1128 		while (dep != NULL) {
1129 			if ((dep->de_hash == hash) &&
1130 			    (namlen == dep->de_namelen) &&
1131 			    bcmp(dep->de_name, name, namlen) == 0) {
1132 				*handle = dep->de_handle;
1133 				mutex_exit(&dcap->dca_lock);
1134 				ncs.ncs_dir_hits.value.ui64++;
1135 				return (DFOUND);
1136 			}
1137 			dep = dep->de_next;
1138 		}
1139 		if (dcp->dc_complete) {
1140 			ret = DNOENT;
1141 		} else {
1142 			ret = DNOCACHE;
1143 		}
1144 		mutex_exit(&dcap->dca_lock);
1145 		return (ret);
1146 	} else {
1147 		mutex_exit(&dcap->dca_lock);
1148 		ncs.ncs_dir_misses.value.ui64++;
1149 		return (DNOCACHE);
1150 	}
1151 }
1152 
1153 /*
1154  * Start a new directory cache. An estimate of the number of
1155  * entries is provided to as a quick check to ensure the directory
1156  * is cacheable.
1157  */
1158 dcret_t
dnlc_dir_start(dcanchor_t * dcap,uint_t num_entries)1159 dnlc_dir_start(dcanchor_t *dcap, uint_t num_entries)
1160 {
1161 	dircache_t *dcp;
1162 
1163 	if (!dnlc_dir_enable ||
1164 	    (num_entries < dnlc_dir_min_size)) {
1165 		return (DNOCACHE);
1166 	}
1167 
1168 	if (num_entries > dnlc_dir_max_size) {
1169 		return (DTOOBIG);
1170 	}
1171 
1172 	mutex_enter(&dc_head.dch_lock);
1173 	mutex_enter(&dcap->dca_lock);
1174 
1175 	if (dcap->dca_dircache == DC_RET_LOW_MEM) {
1176 		dcap->dca_dircache = NULL;
1177 		mutex_exit(&dcap->dca_lock);
1178 		mutex_exit(&dc_head.dch_lock);
1179 		return (DNOMEM);
1180 	}
1181 
1182 	/*
1183 	 * Check if there's currently a cache.
1184 	 * This probably only occurs on a race.
1185 	 */
1186 	if (dcap->dca_dircache != NULL) {
1187 		mutex_exit(&dcap->dca_lock);
1188 		mutex_exit(&dc_head.dch_lock);
1189 		return (DNOCACHE);
1190 	}
1191 
1192 	/*
1193 	 * Allocate the dircache struct, entry and free space hash tables.
1194 	 * These tables are initially just one entry but dynamically resize
1195 	 * when entries and free space are added or removed.
1196 	 */
1197 	if ((dcp = kmem_zalloc(sizeof (dircache_t), KM_NOSLEEP)) == NULL) {
1198 		goto error;
1199 	}
1200 	if ((dcp->dc_namehash = kmem_zalloc(sizeof (dcentry_t *),
1201 	    KM_NOSLEEP)) == NULL) {
1202 		goto error;
1203 	}
1204 	if ((dcp->dc_freehash = kmem_zalloc(sizeof (dcfree_t *),
1205 	    KM_NOSLEEP)) == NULL) {
1206 		goto error;
1207 	}
1208 
1209 	dcp->dc_anchor = dcap; /* set back pointer to anchor */
1210 	dcap->dca_dircache = dcp;
1211 
1212 	/* add into head of global chain */
1213 	dcp->dc_next = dc_head.dch_next;
1214 	dcp->dc_prev = (dircache_t *)&dc_head;
1215 	dcp->dc_next->dc_prev = dcp;
1216 	dc_head.dch_next = dcp;
1217 
1218 	mutex_exit(&dcap->dca_lock);
1219 	mutex_exit(&dc_head.dch_lock);
1220 	ncs.ncs_cur_dirs.value.ui64++;
1221 	ncs.ncs_dirs_cached.value.ui64++;
1222 	return (DOK);
1223 error:
1224 	if (dcp != NULL) {
1225 		if (dcp->dc_namehash) {
1226 			kmem_free(dcp->dc_namehash, sizeof (dcentry_t *));
1227 		}
1228 		kmem_free(dcp, sizeof (dircache_t));
1229 	}
1230 	/*
1231 	 * Must also kmem_free dcp->dc_freehash if more error cases are added
1232 	 */
1233 	mutex_exit(&dcap->dca_lock);
1234 	mutex_exit(&dc_head.dch_lock);
1235 	ncs.ncs_dir_start_nm.value.ui64++;
1236 	return (DNOCACHE);
1237 }
1238 
1239 /*
1240  * Add a directopry entry to a partial or complete directory cache.
1241  */
1242 dcret_t
dnlc_dir_add_entry(dcanchor_t * dcap,const char * name,uint64_t handle)1243 dnlc_dir_add_entry(dcanchor_t *dcap, const char *name, uint64_t handle)
1244 {
1245 	dircache_t *dcp;
1246 	dcentry_t **hp, *dep;
1247 	int hash;
1248 	uint_t capacity;
1249 	uchar_t namlen;
1250 
1251 	/*
1252 	 * Allocate the dcentry struct, including the variable
1253 	 * size name. Note, the null terminator is not copied.
1254 	 *
1255 	 * We do this outside the lock to avoid possible deadlock if
1256 	 * dnlc_dir_reclaim() is called as a result of memory shortage.
1257 	 */
1258 	DNLC_DIR_HASH(name, hash, namlen);
1259 	dep = kmem_alloc(sizeof (dcentry_t) - 1 + namlen, KM_NOSLEEP);
1260 	if (dep == NULL) {
1261 #ifdef DEBUG
1262 		/*
1263 		 * The kmem allocator generates random failures for
1264 		 * KM_NOSLEEP calls (see KMEM_RANDOM_ALLOCATION_FAILURE)
1265 		 * So try again before we blow away a perfectly good cache.
1266 		 * This is done not to cover an error but purely for
1267 		 * performance running a debug kernel.
1268 		 * This random error only occurs in debug mode.
1269 		 */
1270 		dep = kmem_alloc(sizeof (dcentry_t) - 1 + namlen, KM_NOSLEEP);
1271 		if (dep != NULL)
1272 			goto ok;
1273 #endif
1274 		ncs.ncs_dir_add_nm.value.ui64++;
1275 		/*
1276 		 * Free a directory cache. This may be the one we are
1277 		 * called with.
1278 		 */
1279 		dnlc_dir_reclaim(NULL);
1280 		dep = kmem_alloc(sizeof (dcentry_t) - 1 + namlen, KM_NOSLEEP);
1281 		if (dep == NULL) {
1282 			/*
1283 			 * still no memory, better delete this cache
1284 			 */
1285 			mutex_enter(&dcap->dca_lock);
1286 			dcp = (dircache_t *)dcap->dca_dircache;
1287 			if (VALID_DIR_CACHE(dcp)) {
1288 				dnlc_dir_abort(dcp);
1289 				dcap->dca_dircache = DC_RET_LOW_MEM;
1290 			}
1291 			mutex_exit(&dcap->dca_lock);
1292 			ncs.ncs_dir_addabort.value.ui64++;
1293 			return (DNOCACHE);
1294 		}
1295 		/*
1296 		 * fall through as if the 1st kmem_alloc had worked
1297 		 */
1298 	}
1299 #ifdef DEBUG
1300 ok:
1301 #endif
1302 	mutex_enter(&dcap->dca_lock);
1303 	dcp = (dircache_t *)dcap->dca_dircache;
1304 	if (VALID_DIR_CACHE(dcp)) {
1305 		/*
1306 		 * If the total number of entries goes above the max
1307 		 * then free this cache
1308 		 */
1309 		if ((dcp->dc_num_entries + dcp->dc_num_free) >
1310 		    dnlc_dir_max_size) {
1311 			mutex_exit(&dcap->dca_lock);
1312 			dnlc_dir_purge(dcap);
1313 			kmem_free(dep, sizeof (dcentry_t) - 1 + namlen);
1314 			ncs.ncs_dir_add_max.value.ui64++;
1315 			return (DTOOBIG);
1316 		}
1317 		dcp->dc_num_entries++;
1318 		capacity = (dcp->dc_nhash_mask + 1) << dnlc_dir_hash_size_shift;
1319 		if (dcp->dc_num_entries >=
1320 		    (capacity << dnlc_dir_hash_resize_shift)) {
1321 			dnlc_dir_adjust_nhash(dcp);
1322 		}
1323 		hp = &dcp->dc_namehash[hash & dcp->dc_nhash_mask];
1324 
1325 		/*
1326 		 * Initialise and chain in new entry
1327 		 */
1328 		dep->de_handle = handle;
1329 		dep->de_hash = hash;
1330 		/*
1331 		 * Note de_namelen is a uchar_t to conserve space
1332 		 * and alignment padding. The max length of any
1333 		 * pathname component is defined as MAXNAMELEN
1334 		 * which is 256 (including the terminating null).
1335 		 * So provided this doesn't change, we don't include the null,
1336 		 * we always use bcmp to compare strings, and we don't
1337 		 * start storing full names, then we are ok.
1338 		 * The space savings is worth it.
1339 		 */
1340 		dep->de_namelen = namlen;
1341 		bcopy(name, dep->de_name, namlen);
1342 		dep->de_next = *hp;
1343 		*hp = dep;
1344 		dcp->dc_actime = ddi_get_lbolt64();
1345 		mutex_exit(&dcap->dca_lock);
1346 		ncs.ncs_dir_num_ents.value.ui64++;
1347 		return (DOK);
1348 	} else {
1349 		mutex_exit(&dcap->dca_lock);
1350 		kmem_free(dep, sizeof (dcentry_t) - 1 + namlen);
1351 		return (DNOCACHE);
1352 	}
1353 }
1354 
1355 /*
1356  * Add free space to a partial or complete directory cache.
1357  */
1358 dcret_t
dnlc_dir_add_space(dcanchor_t * dcap,uint_t len,uint64_t handle)1359 dnlc_dir_add_space(dcanchor_t *dcap, uint_t len, uint64_t handle)
1360 {
1361 	dircache_t *dcp;
1362 	dcfree_t *dfp, **hp;
1363 	uint_t capacity;
1364 
1365 	/*
1366 	 * We kmem_alloc outside the lock to avoid possible deadlock if
1367 	 * dnlc_dir_reclaim() is called as a result of memory shortage.
1368 	 */
1369 	dfp = kmem_cache_alloc(dnlc_dir_space_cache, KM_NOSLEEP);
1370 	if (dfp == NULL) {
1371 #ifdef DEBUG
1372 		/*
1373 		 * The kmem allocator generates random failures for
1374 		 * KM_NOSLEEP calls (see KMEM_RANDOM_ALLOCATION_FAILURE)
1375 		 * So try again before we blow away a perfectly good cache.
1376 		 * This random error only occurs in debug mode
1377 		 */
1378 		dfp = kmem_cache_alloc(dnlc_dir_space_cache, KM_NOSLEEP);
1379 		if (dfp != NULL)
1380 			goto ok;
1381 #endif
1382 		ncs.ncs_dir_add_nm.value.ui64++;
1383 		/*
1384 		 * Free a directory cache. This may be the one we are
1385 		 * called with.
1386 		 */
1387 		dnlc_dir_reclaim(NULL);
1388 		dfp = kmem_cache_alloc(dnlc_dir_space_cache, KM_NOSLEEP);
1389 		if (dfp == NULL) {
1390 			/*
1391 			 * still no memory, better delete this cache
1392 			 */
1393 			mutex_enter(&dcap->dca_lock);
1394 			dcp = (dircache_t *)dcap->dca_dircache;
1395 			if (VALID_DIR_CACHE(dcp)) {
1396 				dnlc_dir_abort(dcp);
1397 				dcap->dca_dircache = DC_RET_LOW_MEM;
1398 			}
1399 			mutex_exit(&dcap->dca_lock);
1400 			ncs.ncs_dir_addabort.value.ui64++;
1401 			return (DNOCACHE);
1402 		}
1403 		/*
1404 		 * fall through as if the 1st kmem_alloc had worked
1405 		 */
1406 	}
1407 
1408 #ifdef DEBUG
1409 ok:
1410 #endif
1411 	mutex_enter(&dcap->dca_lock);
1412 	dcp = (dircache_t *)dcap->dca_dircache;
1413 	if (VALID_DIR_CACHE(dcp)) {
1414 		if ((dcp->dc_num_entries + dcp->dc_num_free) >
1415 		    dnlc_dir_max_size) {
1416 			mutex_exit(&dcap->dca_lock);
1417 			dnlc_dir_purge(dcap);
1418 			kmem_cache_free(dnlc_dir_space_cache, dfp);
1419 			ncs.ncs_dir_add_max.value.ui64++;
1420 			return (DTOOBIG);
1421 		}
1422 		dcp->dc_num_free++;
1423 		capacity = (dcp->dc_fhash_mask + 1) << dnlc_dir_hash_size_shift;
1424 		if (dcp->dc_num_free >=
1425 		    (capacity << dnlc_dir_hash_resize_shift)) {
1426 			dnlc_dir_adjust_fhash(dcp);
1427 		}
1428 		/*
1429 		 * Initialise and chain a new entry
1430 		 */
1431 		dfp->df_handle = handle;
1432 		dfp->df_len = len;
1433 		dcp->dc_actime = ddi_get_lbolt64();
1434 		hp = &(dcp->dc_freehash[DDFHASH(handle, dcp)]);
1435 		dfp->df_next = *hp;
1436 		*hp = dfp;
1437 		mutex_exit(&dcap->dca_lock);
1438 		ncs.ncs_dir_num_ents.value.ui64++;
1439 		return (DOK);
1440 	} else {
1441 		mutex_exit(&dcap->dca_lock);
1442 		kmem_cache_free(dnlc_dir_space_cache, dfp);
1443 		return (DNOCACHE);
1444 	}
1445 }
1446 
1447 /*
1448  * Mark a directory cache as complete.
1449  */
1450 void
dnlc_dir_complete(dcanchor_t * dcap)1451 dnlc_dir_complete(dcanchor_t *dcap)
1452 {
1453 	dircache_t *dcp;
1454 
1455 	mutex_enter(&dcap->dca_lock);
1456 	dcp = (dircache_t *)dcap->dca_dircache;
1457 	if (VALID_DIR_CACHE(dcp)) {
1458 		dcp->dc_complete = B_TRUE;
1459 	}
1460 	mutex_exit(&dcap->dca_lock);
1461 }
1462 
1463 /*
1464  * Internal routine to delete a partial or full directory cache.
1465  * No additional locking needed.
1466  */
1467 static void
dnlc_dir_abort(dircache_t * dcp)1468 dnlc_dir_abort(dircache_t *dcp)
1469 {
1470 	dcentry_t *dep, *nhp;
1471 	dcfree_t *fep, *fhp;
1472 	uint_t nhtsize = dcp->dc_nhash_mask + 1; /* name hash table size */
1473 	uint_t fhtsize = dcp->dc_fhash_mask + 1; /* free hash table size */
1474 	uint_t i;
1475 
1476 	/*
1477 	 * Free up the cached name entries and hash table
1478 	 */
1479 	for (i = 0; i < nhtsize; i++) { /* for each hash bucket */
1480 		nhp = dcp->dc_namehash[i];
1481 		while (nhp != NULL) { /* for each chained entry */
1482 			dep = nhp->de_next;
1483 			kmem_free(nhp, sizeof (dcentry_t) - 1 +
1484 			    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, sizeof (dcentry_t) - 1 +
1581 				    te->de_namelen);
1582 
1583 				/*
1584 				 * If the total number of entries
1585 				 * falls below half the minimum number
1586 				 * of entries then free this cache.
1587 				 */
1588 				if (--dcp->dc_num_entries <
1589 				    (dnlc_dir_min_size >> 1)) {
1590 					mutex_exit(&dcap->dca_lock);
1591 					dnlc_dir_purge(dcap);
1592 				} else {
1593 					mutex_exit(&dcap->dca_lock);
1594 				}
1595 				ncs.ncs_dir_num_ents.value.ui64--;
1596 				return (DFOUND);
1597 			}
1598 			prevpp = &((*prevpp)->de_next);
1599 		}
1600 		if (dcp->dc_complete) {
1601 			ncs.ncs_dir_reme_fai.value.ui64++;
1602 			ret = DNOENT;
1603 		} else {
1604 			ret = DNOCACHE;
1605 		}
1606 		mutex_exit(&dcap->dca_lock);
1607 		return (ret);
1608 	} else {
1609 		mutex_exit(&dcap->dca_lock);
1610 		return (DNOCACHE);
1611 	}
1612 }
1613 
1614 
1615 /*
1616  * Remove free space of at least the given length from a complete
1617  * or partial directory cache.
1618  */
1619 dcret_t
dnlc_dir_rem_space_by_len(dcanchor_t * dcap,uint_t len,uint64_t * handlep)1620 dnlc_dir_rem_space_by_len(dcanchor_t *dcap, uint_t len, uint64_t *handlep)
1621 {
1622 	dircache_t *dcp;
1623 	dcfree_t **prevpp, *tfp;
1624 	uint_t fhtsize; /* free hash table size */
1625 	uint_t i;
1626 	uint_t capacity;
1627 	int ret;
1628 
1629 	if (!dnlc_dir_enable) {
1630 		return (DNOCACHE);
1631 	}
1632 
1633 	mutex_enter(&dcap->dca_lock);
1634 	dcp = (dircache_t *)dcap->dca_dircache;
1635 	if (VALID_DIR_CACHE(dcp)) {
1636 		dcp->dc_actime = ddi_get_lbolt64();
1637 		if (dcp->dc_fhash_mask > 0) { /* ie not minimum */
1638 			capacity = (dcp->dc_fhash_mask + 1) <<
1639 			    dnlc_dir_hash_size_shift;
1640 			if (dcp->dc_num_free <=
1641 			    (capacity >> dnlc_dir_hash_resize_shift)) {
1642 				dnlc_dir_adjust_fhash(dcp);
1643 			}
1644 		}
1645 		/*
1646 		 * Search for an entry of the appropriate size
1647 		 * on a first fit basis.
1648 		 */
1649 		fhtsize = dcp->dc_fhash_mask + 1;
1650 		for (i = 0; i < fhtsize; i++) { /* for each hash bucket */
1651 			prevpp = &(dcp->dc_freehash[i]);
1652 			while (*prevpp != NULL) {
1653 				if ((*prevpp)->df_len >= len) {
1654 					*handlep = (*prevpp)->df_handle;
1655 					tfp = *prevpp;
1656 					*prevpp = (*prevpp)->df_next;
1657 					dcp->dc_num_free--;
1658 					mutex_exit(&dcap->dca_lock);
1659 					kmem_cache_free(dnlc_dir_space_cache,
1660 					    tfp);
1661 					ncs.ncs_dir_num_ents.value.ui64--;
1662 					return (DFOUND);
1663 				}
1664 				prevpp = &((*prevpp)->df_next);
1665 			}
1666 		}
1667 		if (dcp->dc_complete) {
1668 			ret = DNOENT;
1669 		} else {
1670 			ret = DNOCACHE;
1671 		}
1672 		mutex_exit(&dcap->dca_lock);
1673 		return (ret);
1674 	} else {
1675 		mutex_exit(&dcap->dca_lock);
1676 		return (DNOCACHE);
1677 	}
1678 }
1679 
1680 /*
1681  * Remove free space with the given handle from a complete or partial
1682  * directory cache.
1683  */
1684 dcret_t
dnlc_dir_rem_space_by_handle(dcanchor_t * dcap,uint64_t handle)1685 dnlc_dir_rem_space_by_handle(dcanchor_t *dcap, uint64_t handle)
1686 {
1687 	dircache_t *dcp;
1688 	dcfree_t **prevpp, *tfp;
1689 	uint_t capacity;
1690 	int ret;
1691 
1692 	if (!dnlc_dir_enable) {
1693 		return (DNOCACHE);
1694 	}
1695 
1696 	mutex_enter(&dcap->dca_lock);
1697 	dcp = (dircache_t *)dcap->dca_dircache;
1698 	if (VALID_DIR_CACHE(dcp)) {
1699 		dcp->dc_actime = ddi_get_lbolt64();
1700 		if (dcp->dc_fhash_mask > 0) { /* ie not minimum */
1701 			capacity = (dcp->dc_fhash_mask + 1) <<
1702 			    dnlc_dir_hash_size_shift;
1703 			if (dcp->dc_num_free <=
1704 			    (capacity >> dnlc_dir_hash_resize_shift)) {
1705 				dnlc_dir_adjust_fhash(dcp);
1706 			}
1707 		}
1708 
1709 		/*
1710 		 * search for the exact entry
1711 		 */
1712 		prevpp = &(dcp->dc_freehash[DDFHASH(handle, dcp)]);
1713 		while (*prevpp != NULL) {
1714 			if ((*prevpp)->df_handle == handle) {
1715 				tfp = *prevpp;
1716 				*prevpp = (*prevpp)->df_next;
1717 				dcp->dc_num_free--;
1718 				mutex_exit(&dcap->dca_lock);
1719 				kmem_cache_free(dnlc_dir_space_cache, tfp);
1720 				ncs.ncs_dir_num_ents.value.ui64--;
1721 				return (DFOUND);
1722 			}
1723 			prevpp = &((*prevpp)->df_next);
1724 		}
1725 		if (dcp->dc_complete) {
1726 			ncs.ncs_dir_rems_fai.value.ui64++;
1727 			ret = DNOENT;
1728 		} else {
1729 			ret = DNOCACHE;
1730 		}
1731 		mutex_exit(&dcap->dca_lock);
1732 		return (ret);
1733 	} else {
1734 		mutex_exit(&dcap->dca_lock);
1735 		return (DNOCACHE);
1736 	}
1737 }
1738 
1739 /*
1740  * Update the handle of an directory cache entry.
1741  */
1742 dcret_t
dnlc_dir_update(dcanchor_t * dcap,const char * name,uint64_t handle)1743 dnlc_dir_update(dcanchor_t *dcap, const char *name, uint64_t handle)
1744 {
1745 	dircache_t *dcp;
1746 	dcentry_t *dep;
1747 	int hash;
1748 	int ret;
1749 	uchar_t namlen;
1750 
1751 	if (!dnlc_dir_enable) {
1752 		return (DNOCACHE);
1753 	}
1754 
1755 	mutex_enter(&dcap->dca_lock);
1756 	dcp = (dircache_t *)dcap->dca_dircache;
1757 	if (VALID_DIR_CACHE(dcp)) {
1758 		dcp->dc_actime = ddi_get_lbolt64();
1759 		DNLC_DIR_HASH(name, hash, namlen);
1760 		dep = dcp->dc_namehash[hash & dcp->dc_nhash_mask];
1761 		while (dep != NULL) {
1762 			if ((dep->de_hash == hash) &&
1763 			    (namlen == dep->de_namelen) &&
1764 			    bcmp(dep->de_name, name, namlen) == 0) {
1765 				dep->de_handle = handle;
1766 				mutex_exit(&dcap->dca_lock);
1767 				return (DFOUND);
1768 			}
1769 			dep = dep->de_next;
1770 		}
1771 		if (dcp->dc_complete) {
1772 			ncs.ncs_dir_upd_fail.value.ui64++;
1773 			ret = DNOENT;
1774 		} else {
1775 			ret = DNOCACHE;
1776 		}
1777 		mutex_exit(&dcap->dca_lock);
1778 		return (ret);
1779 	} else {
1780 		mutex_exit(&dcap->dca_lock);
1781 		return (DNOCACHE);
1782 	}
1783 }
1784 
1785 void
dnlc_dir_fini(dcanchor_t * dcap)1786 dnlc_dir_fini(dcanchor_t *dcap)
1787 {
1788 	dircache_t *dcp;
1789 
1790 	mutex_enter(&dc_head.dch_lock);
1791 	mutex_enter(&dcap->dca_lock);
1792 	dcp = (dircache_t *)dcap->dca_dircache;
1793 	if (VALID_DIR_CACHE(dcp)) {
1794 		/*
1795 		 * Unchain from global list
1796 		 */
1797 		ncs.ncs_dir_finipurg.value.ui64++;
1798 		dcp->dc_prev->dc_next = dcp->dc_next;
1799 		dcp->dc_next->dc_prev = dcp->dc_prev;
1800 	} else {
1801 		dcp = NULL;
1802 	}
1803 	dcap->dca_dircache = NULL;
1804 	mutex_exit(&dcap->dca_lock);
1805 	mutex_exit(&dc_head.dch_lock);
1806 	mutex_destroy(&dcap->dca_lock);
1807 	if (dcp) {
1808 		dnlc_dir_abort(dcp);
1809 	}
1810 }
1811 
1812 /*
1813  * Reclaim callback for dnlc directory caching.
1814  * Invoked by the kernel memory allocator when memory gets tight.
1815  * This is a pretty serious condition and can lead easily lead to system
1816  * hangs if not enough space is returned.
1817  *
1818  * Deciding which directory (or directories) to purge is tricky.
1819  * Purging everything is an overkill, but purging just the oldest used
1820  * was found to lead to hangs. The largest cached directories use the
1821  * most memory, but take the most effort to rebuild, whereas the smaller
1822  * ones have little value and give back little space. So what to do?
1823  *
1824  * The current policy is to continue purging the oldest used directories
1825  * until at least dnlc_dir_min_reclaim directory entries have been purged.
1826  */
1827 /*ARGSUSED*/
1828 static void
dnlc_dir_reclaim(void * unused)1829 dnlc_dir_reclaim(void *unused)
1830 {
1831 	dircache_t *dcp, *oldest;
1832 	uint_t dirent_cnt = 0;
1833 
1834 	mutex_enter(&dc_head.dch_lock);
1835 	while (dirent_cnt < dnlc_dir_min_reclaim) {
1836 		dcp = dc_head.dch_next;
1837 		oldest = NULL;
1838 		while (dcp != (dircache_t *)&dc_head) {
1839 			if (oldest == NULL) {
1840 				oldest = dcp;
1841 			} else {
1842 				if (dcp->dc_actime < oldest->dc_actime) {
1843 					oldest = dcp;
1844 				}
1845 			}
1846 			dcp = dcp->dc_next;
1847 		}
1848 		if (oldest == NULL) {
1849 			/* nothing to delete */
1850 			mutex_exit(&dc_head.dch_lock);
1851 			return;
1852 		}
1853 		/*
1854 		 * remove from directory chain and purge
1855 		 */
1856 		oldest->dc_prev->dc_next = oldest->dc_next;
1857 		oldest->dc_next->dc_prev = oldest->dc_prev;
1858 		mutex_enter(&oldest->dc_anchor->dca_lock);
1859 		/*
1860 		 * If this was the last entry then it must be too large.
1861 		 * Mark it as such by saving a special dircache_t
1862 		 * pointer (DC_RET_LOW_MEM) in the anchor. The error DNOMEM
1863 		 * will be presented to the caller of dnlc_dir_start()
1864 		 */
1865 		if (oldest->dc_next == oldest->dc_prev) {
1866 			oldest->dc_anchor->dca_dircache = DC_RET_LOW_MEM;
1867 			ncs.ncs_dir_rec_last.value.ui64++;
1868 		} else {
1869 			oldest->dc_anchor->dca_dircache = NULL;
1870 			ncs.ncs_dir_recl_any.value.ui64++;
1871 		}
1872 		mutex_exit(&oldest->dc_anchor->dca_lock);
1873 		dirent_cnt += oldest->dc_num_entries;
1874 		dnlc_dir_abort(oldest);
1875 	}
1876 	mutex_exit(&dc_head.dch_lock);
1877 }
1878 
1879 /*
1880  * Dynamically grow or shrink the size of the name hash table
1881  */
1882 static void
dnlc_dir_adjust_nhash(dircache_t * dcp)1883 dnlc_dir_adjust_nhash(dircache_t *dcp)
1884 {
1885 	dcentry_t **newhash, *dep, **nhp, *tep;
1886 	uint_t newsize;
1887 	uint_t oldsize;
1888 	uint_t newsizemask;
1889 	int i;
1890 
1891 	/*
1892 	 * Allocate new hash table
1893 	 */
1894 	newsize = dcp->dc_num_entries >> dnlc_dir_hash_size_shift;
1895 	newhash = kmem_zalloc(sizeof (dcentry_t *) * newsize, KM_NOSLEEP);
1896 	if (newhash == NULL) {
1897 		/*
1898 		 * System is short on memory just return
1899 		 * Note, the old hash table is still usable.
1900 		 * This return is unlikely to repeatedy occur, because
1901 		 * either some other directory caches will be reclaimed
1902 		 * due to memory shortage, thus freeing memory, or this
1903 		 * directory cahe will be reclaimed.
1904 		 */
1905 		return;
1906 	}
1907 	oldsize = dcp->dc_nhash_mask + 1;
1908 	dcp->dc_nhash_mask = newsizemask = newsize - 1;
1909 
1910 	/*
1911 	 * Move entries from the old table to the new
1912 	 */
1913 	for (i = 0; i < oldsize; i++) { /* for each hash bucket */
1914 		dep = dcp->dc_namehash[i];
1915 		while (dep != NULL) { /* for each chained entry */
1916 			tep = dep;
1917 			dep = dep->de_next;
1918 			nhp = &newhash[tep->de_hash & newsizemask];
1919 			tep->de_next = *nhp;
1920 			*nhp = tep;
1921 		}
1922 	}
1923 
1924 	/*
1925 	 * delete old hash table and set new one in place
1926 	 */
1927 	kmem_free(dcp->dc_namehash, sizeof (dcentry_t *) * oldsize);
1928 	dcp->dc_namehash = newhash;
1929 }
1930 
1931 /*
1932  * Dynamically grow or shrink the size of the free space hash table
1933  */
1934 static void
dnlc_dir_adjust_fhash(dircache_t * dcp)1935 dnlc_dir_adjust_fhash(dircache_t *dcp)
1936 {
1937 	dcfree_t **newhash, *dfp, **nhp, *tfp;
1938 	uint_t newsize;
1939 	uint_t oldsize;
1940 	int i;
1941 
1942 	/*
1943 	 * Allocate new hash table
1944 	 */
1945 	newsize = dcp->dc_num_free >> dnlc_dir_hash_size_shift;
1946 	newhash = kmem_zalloc(sizeof (dcfree_t *) * newsize, KM_NOSLEEP);
1947 	if (newhash == NULL) {
1948 		/*
1949 		 * System is short on memory just return
1950 		 * Note, the old hash table is still usable.
1951 		 * This return is unlikely to repeatedy occur, because
1952 		 * either some other directory caches will be reclaimed
1953 		 * due to memory shortage, thus freeing memory, or this
1954 		 * directory cahe will be reclaimed.
1955 		 */
1956 		return;
1957 	}
1958 	oldsize = dcp->dc_fhash_mask + 1;
1959 	dcp->dc_fhash_mask = newsize - 1;
1960 
1961 	/*
1962 	 * Move entries from the old table to the new
1963 	 */
1964 	for (i = 0; i < oldsize; i++) { /* for each hash bucket */
1965 		dfp = dcp->dc_freehash[i];
1966 		while (dfp != NULL) { /* for each chained entry */
1967 			tfp = dfp;
1968 			dfp = dfp->df_next;
1969 			nhp = &newhash[DDFHASH(tfp->df_handle, dcp)];
1970 			tfp->df_next = *nhp;
1971 			*nhp = tfp;
1972 		}
1973 	}
1974 
1975 	/*
1976 	 * delete old hash table and set new one in place
1977 	 */
1978 	kmem_free(dcp->dc_freehash, sizeof (dcfree_t *) * oldsize);
1979 	dcp->dc_freehash = newhash;
1980 }
1981