xref: /freebsd/sys/kern/vfs_cache.c (revision a0409676120c1e558d0ade943019934e0f15118d)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993, 1995
5  *	The Regents of the University of California.  All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * Poul-Henning Kamp of the FreeBSD Project.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  * 3. Neither the name of the University nor the names of its contributors
19  *    may be used to endorse or promote products derived from this software
20  *    without specific prior written permission.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32  * SUCH DAMAGE.
33  *
34  *	@(#)vfs_cache.c	8.5 (Berkeley) 3/22/95
35  */
36 
37 #include <sys/cdefs.h>
38 __FBSDID("$FreeBSD$");
39 
40 #include "opt_ddb.h"
41 #include "opt_ktrace.h"
42 
43 #include <sys/param.h>
44 #include <sys/systm.h>
45 #include <sys/capsicum.h>
46 #include <sys/counter.h>
47 #include <sys/filedesc.h>
48 #include <sys/fnv_hash.h>
49 #include <sys/kernel.h>
50 #include <sys/ktr.h>
51 #include <sys/lock.h>
52 #include <sys/malloc.h>
53 #include <sys/fcntl.h>
54 #include <sys/jail.h>
55 #include <sys/mount.h>
56 #include <sys/namei.h>
57 #include <sys/proc.h>
58 #include <sys/seqc.h>
59 #include <sys/sdt.h>
60 #include <sys/smr.h>
61 #include <sys/smp.h>
62 #include <sys/syscallsubr.h>
63 #include <sys/sysctl.h>
64 #include <sys/sysproto.h>
65 #include <sys/vnode.h>
66 #include <ck_queue.h>
67 #ifdef KTRACE
68 #include <sys/ktrace.h>
69 #endif
70 #ifdef INVARIANTS
71 #include <machine/_inttypes.h>
72 #endif
73 
74 #include <sys/capsicum.h>
75 
76 #include <security/audit/audit.h>
77 #include <security/mac/mac_framework.h>
78 
79 #ifdef DDB
80 #include <ddb/ddb.h>
81 #endif
82 
83 #include <vm/uma.h>
84 
85 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
86     "Name cache");
87 
88 SDT_PROVIDER_DECLARE(vfs);
89 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
90     "struct vnode *");
91 SDT_PROBE_DEFINE3(vfs, namecache, enter, duplicate, "struct vnode *", "char *",
92     "struct vnode *");
93 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
94     "char *");
95 SDT_PROBE_DEFINE2(vfs, namecache, fullpath_smr, hit, "struct vnode *",
96     "const char *");
97 SDT_PROBE_DEFINE4(vfs, namecache, fullpath_smr, miss, "struct vnode *",
98     "struct namecache *", "int", "int");
99 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
100 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
101     "char *", "struct vnode *");
102 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
103 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
104     "struct vnode *", "char *");
105 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
106     "struct vnode *");
107 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
108     "struct vnode *", "char *");
109 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
110     "char *");
111 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, hit, "struct vnode *",
112     "struct componentname *");
113 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, miss, "struct vnode *",
114     "struct componentname *");
115 SDT_PROBE_DEFINE1(vfs, namecache, purge, done, "struct vnode *");
116 SDT_PROBE_DEFINE1(vfs, namecache, purge, batch, "int");
117 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
118 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
119 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
120     "struct vnode *");
121 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *",
122     "char *");
123 SDT_PROBE_DEFINE2(vfs, namecache, evict_negative, done, "struct vnode *",
124     "char *");
125 SDT_PROBE_DEFINE1(vfs, namecache, symlink, alloc__fail, "size_t");
126 
127 SDT_PROBE_DEFINE3(vfs, fplookup, lookup, done, "struct nameidata", "int", "bool");
128 SDT_PROBE_DECLARE(vfs, namei, lookup, entry);
129 SDT_PROBE_DECLARE(vfs, namei, lookup, return);
130 
131 /*
132  * This structure describes the elements in the cache of recent
133  * names looked up by namei.
134  */
135 struct negstate {
136 	u_char neg_flag;
137 	u_char neg_hit;
138 };
139 _Static_assert(sizeof(struct negstate) <= sizeof(struct vnode *),
140     "the state must fit in a union with a pointer without growing it");
141 
142 struct	namecache {
143 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
144 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
145 	CK_SLIST_ENTRY(namecache) nc_hash;/* hash chain */
146 	struct	vnode *nc_dvp;		/* vnode of parent of name */
147 	union {
148 		struct	vnode *nu_vp;	/* vnode the name refers to */
149 		struct	negstate nu_neg;/* negative entry state */
150 	} n_un;
151 	u_char	nc_flag;		/* flag bits */
152 	u_char	nc_nlen;		/* length of name */
153 	char	nc_name[0];		/* segment name + nul */
154 };
155 
156 /*
157  * struct namecache_ts repeats struct namecache layout up to the
158  * nc_nlen member.
159  * struct namecache_ts is used in place of struct namecache when time(s) need
160  * to be stored.  The nc_dotdottime field is used when a cache entry is mapping
161  * both a non-dotdot directory name plus dotdot for the directory's
162  * parent.
163  *
164  * See below for alignment requirement.
165  */
166 struct	namecache_ts {
167 	struct	timespec nc_time;	/* timespec provided by fs */
168 	struct	timespec nc_dotdottime;	/* dotdot timespec provided by fs */
169 	int	nc_ticks;		/* ticks value when entry was added */
170 	int	nc_pad;
171 	struct namecache nc_nc;
172 };
173 
174 TAILQ_HEAD(cache_freebatch, namecache);
175 
176 /*
177  * At least mips n32 performs 64-bit accesses to timespec as found
178  * in namecache_ts and requires them to be aligned. Since others
179  * may be in the same spot suffer a little bit and enforce the
180  * alignment for everyone. Note this is a nop for 64-bit platforms.
181  */
182 #define CACHE_ZONE_ALIGNMENT	UMA_ALIGNOF(time_t)
183 
184 /*
185  * TODO: the initial value of CACHE_PATH_CUTOFF was inherited from the
186  * 4.4 BSD codebase. Later on struct namecache was tweaked to become
187  * smaller and the value was bumped to retain the total size, but it
188  * was never re-evaluated for suitability. A simple test counting
189  * lengths during package building shows that the value of 45 covers
190  * about 86% of all added entries, reaching 99% at 65.
191  *
192  * Regardless of the above, use of dedicated zones instead of malloc may be
193  * inducing additional waste. This may be hard to address as said zones are
194  * tied to VFS SMR. Even if retaining them, the current split should be
195  * re-evaluated.
196  */
197 #ifdef __LP64__
198 #define	CACHE_PATH_CUTOFF	45
199 #define	CACHE_LARGE_PAD		6
200 #else
201 #define	CACHE_PATH_CUTOFF	41
202 #define	CACHE_LARGE_PAD		2
203 #endif
204 
205 #define CACHE_ZONE_SMALL_SIZE		(offsetof(struct namecache, nc_name) + CACHE_PATH_CUTOFF + 1)
206 #define CACHE_ZONE_SMALL_TS_SIZE	(offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_SMALL_SIZE)
207 #define CACHE_ZONE_LARGE_SIZE		(offsetof(struct namecache, nc_name) + NAME_MAX + 1 + CACHE_LARGE_PAD)
208 #define CACHE_ZONE_LARGE_TS_SIZE	(offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_LARGE_SIZE)
209 
210 _Static_assert((CACHE_ZONE_SMALL_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
211 _Static_assert((CACHE_ZONE_SMALL_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
212 _Static_assert((CACHE_ZONE_LARGE_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
213 _Static_assert((CACHE_ZONE_LARGE_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
214 
215 #define	nc_vp		n_un.nu_vp
216 #define	nc_neg		n_un.nu_neg
217 
218 /*
219  * Flags in namecache.nc_flag
220  */
221 #define NCF_WHITE	0x01
222 #define NCF_ISDOTDOT	0x02
223 #define	NCF_TS		0x04
224 #define	NCF_DTS		0x08
225 #define	NCF_DVDROP	0x10
226 #define	NCF_NEGATIVE	0x20
227 #define	NCF_INVALID	0x40
228 #define	NCF_WIP		0x80
229 
230 /*
231  * Flags in negstate.neg_flag
232  */
233 #define NEG_HOT		0x01
234 
235 static bool	cache_neg_evict_cond(u_long lnumcache);
236 
237 /*
238  * Mark an entry as invalid.
239  *
240  * This is called before it starts getting deconstructed.
241  */
242 static void
243 cache_ncp_invalidate(struct namecache *ncp)
244 {
245 
246 	KASSERT((ncp->nc_flag & NCF_INVALID) == 0,
247 	    ("%s: entry %p already invalid", __func__, ncp));
248 	atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_INVALID);
249 	atomic_thread_fence_rel();
250 }
251 
252 /*
253  * Check whether the entry can be safely used.
254  *
255  * All places which elide locks are supposed to call this after they are
256  * done with reading from an entry.
257  */
258 #define cache_ncp_canuse(ncp)	({					\
259 	struct namecache *_ncp = (ncp);					\
260 	u_char _nc_flag;						\
261 									\
262 	atomic_thread_fence_acq();					\
263 	_nc_flag = atomic_load_char(&_ncp->nc_flag);			\
264 	__predict_true((_nc_flag & (NCF_INVALID | NCF_WIP)) == 0);	\
265 })
266 
267 /*
268  * Like the above but also checks NCF_WHITE.
269  */
270 #define cache_fpl_neg_ncp_canuse(ncp)	({				\
271 	struct namecache *_ncp = (ncp);					\
272 	u_char _nc_flag;						\
273 									\
274 	atomic_thread_fence_acq();					\
275 	_nc_flag = atomic_load_char(&_ncp->nc_flag);			\
276 	__predict_true((_nc_flag & (NCF_INVALID | NCF_WIP | NCF_WHITE)) == 0);	\
277 })
278 
279 VFS_SMR_DECLARE;
280 
281 static SYSCTL_NODE(_vfs_cache, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
282     "Name cache parameters");
283 
284 static u_int __read_mostly	ncsize; /* the size as computed on creation or resizing */
285 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, size, CTLFLAG_RW, &ncsize, 0,
286     "Total namecache capacity");
287 
288 u_int ncsizefactor = 2;
289 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, sizefactor, CTLFLAG_RW, &ncsizefactor, 0,
290     "Size factor for namecache");
291 
292 static u_long __read_mostly	ncnegfactor = 5; /* ratio of negative entries */
293 SYSCTL_ULONG(_vfs_cache_param, OID_AUTO, negfactor, CTLFLAG_RW, &ncnegfactor, 0,
294     "Ratio of negative namecache entries");
295 
296 /*
297  * Negative entry % of namecache capacity above which automatic eviction is allowed.
298  *
299  * Check cache_neg_evict_cond for details.
300  */
301 static u_int ncnegminpct = 3;
302 
303 static u_int __read_mostly     neg_min; /* the above recomputed against ncsize */
304 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, negmin, CTLFLAG_RD, &neg_min, 0,
305     "Negative entry count above which automatic eviction is allowed");
306 
307 /*
308  * Structures associated with name caching.
309  */
310 #define NCHHASH(hash) \
311 	(&nchashtbl[(hash) & nchash])
312 static __read_mostly CK_SLIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
313 static u_long __read_mostly	nchash;			/* size of hash table */
314 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
315     "Size of namecache hash table");
316 static u_long __exclusive_cache_line	numneg;	/* number of negative entries allocated */
317 static u_long __exclusive_cache_line	numcache;/* number of cache entries allocated */
318 
319 struct nchstats	nchstats;		/* cache effectiveness statistics */
320 
321 static bool __read_frequently cache_fast_revlookup = true;
322 SYSCTL_BOOL(_vfs, OID_AUTO, cache_fast_revlookup, CTLFLAG_RW,
323     &cache_fast_revlookup, 0, "");
324 
325 static bool __read_mostly cache_rename_add = true;
326 SYSCTL_BOOL(_vfs, OID_AUTO, cache_rename_add, CTLFLAG_RW,
327     &cache_rename_add, 0, "");
328 
329 static u_int __exclusive_cache_line neg_cycle;
330 
331 #define ncneghash	3
332 #define	numneglists	(ncneghash + 1)
333 
334 struct neglist {
335 	struct mtx		nl_evict_lock;
336 	struct mtx		nl_lock __aligned(CACHE_LINE_SIZE);
337 	TAILQ_HEAD(, namecache) nl_list;
338 	TAILQ_HEAD(, namecache) nl_hotlist;
339 	u_long			nl_hotnum;
340 } __aligned(CACHE_LINE_SIZE);
341 
342 static struct neglist neglists[numneglists];
343 
344 static inline struct neglist *
345 NCP2NEGLIST(struct namecache *ncp)
346 {
347 
348 	return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
349 }
350 
351 static inline struct negstate *
352 NCP2NEGSTATE(struct namecache *ncp)
353 {
354 
355 	MPASS(atomic_load_char(&ncp->nc_flag) & NCF_NEGATIVE);
356 	return (&ncp->nc_neg);
357 }
358 
359 #define	numbucketlocks (ncbuckethash + 1)
360 static u_int __read_mostly  ncbuckethash;
361 static struct mtx_padalign __read_mostly  *bucketlocks;
362 #define	HASH2BUCKETLOCK(hash) \
363 	((struct mtx *)(&bucketlocks[((hash) & ncbuckethash)]))
364 
365 #define	numvnodelocks (ncvnodehash + 1)
366 static u_int __read_mostly  ncvnodehash;
367 static struct mtx __read_mostly *vnodelocks;
368 static inline struct mtx *
369 VP2VNODELOCK(struct vnode *vp)
370 {
371 
372 	return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
373 }
374 
375 static void
376 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
377 {
378 	struct namecache_ts *ncp_ts;
379 
380 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
381 	    (tsp == NULL && ticksp == NULL),
382 	    ("No NCF_TS"));
383 
384 	if (tsp == NULL)
385 		return;
386 
387 	ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
388 	*tsp = ncp_ts->nc_time;
389 	*ticksp = ncp_ts->nc_ticks;
390 }
391 
392 #ifdef DEBUG_CACHE
393 static int __read_mostly	doingcache = 1;	/* 1 => enable the cache */
394 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
395     "VFS namecache enabled");
396 #endif
397 
398 /* Export size information to userland */
399 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
400     sizeof(struct namecache), "sizeof(struct namecache)");
401 
402 /*
403  * The new name cache statistics
404  */
405 static SYSCTL_NODE(_vfs_cache, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
406     "Name cache statistics");
407 
408 #define STATNODE_ULONG(name, varname, descr)					\
409 	SYSCTL_ULONG(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
410 #define STATNODE_COUNTER(name, varname, descr)					\
411 	static COUNTER_U64_DEFINE_EARLY(varname);				\
412 	SYSCTL_COUNTER_U64(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, \
413 	    descr);
414 STATNODE_ULONG(neg, numneg, "Number of negative cache entries");
415 STATNODE_ULONG(count, numcache, "Number of cache entries");
416 STATNODE_COUNTER(heldvnodes, numcachehv, "Number of namecache entries with vnodes held");
417 STATNODE_COUNTER(drops, numdrops, "Number of dropped entries due to reaching the limit");
418 STATNODE_COUNTER(dothits, dothits, "Number of '.' hits");
419 STATNODE_COUNTER(dotdothis, dotdothits, "Number of '..' hits");
420 STATNODE_COUNTER(miss, nummiss, "Number of cache misses");
421 STATNODE_COUNTER(misszap, nummisszap, "Number of cache misses we do not want to cache");
422 STATNODE_COUNTER(posszaps, numposzaps,
423     "Number of cache hits (positive) we do not want to cache");
424 STATNODE_COUNTER(poshits, numposhits, "Number of cache hits (positive)");
425 STATNODE_COUNTER(negzaps, numnegzaps,
426     "Number of cache hits (negative) we do not want to cache");
427 STATNODE_COUNTER(neghits, numneghits, "Number of cache hits (negative)");
428 /* These count for vn_getcwd(), too. */
429 STATNODE_COUNTER(fullpathcalls, numfullpathcalls, "Number of fullpath search calls");
430 STATNODE_COUNTER(fullpathfail1, numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
431 STATNODE_COUNTER(fullpathfail2, numfullpathfail2,
432     "Number of fullpath search errors (VOP_VPTOCNP failures)");
433 STATNODE_COUNTER(fullpathfail4, numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
434 STATNODE_COUNTER(fullpathfound, numfullpathfound, "Number of successful fullpath calls");
435 STATNODE_COUNTER(symlinktoobig, symlinktoobig, "Number of times symlink did not fit the cache");
436 
437 /*
438  * Debug or developer statistics.
439  */
440 static SYSCTL_NODE(_vfs_cache, OID_AUTO, debug, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
441     "Name cache debugging");
442 #define DEBUGNODE_ULONG(name, varname, descr)					\
443 	SYSCTL_ULONG(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
444 #define DEBUGNODE_COUNTER(name, varname, descr)					\
445 	static COUNTER_U64_DEFINE_EARLY(varname);				\
446 	SYSCTL_COUNTER_U64(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, \
447 	    descr);
448 DEBUGNODE_COUNTER(zap_bucket_relock_success, zap_bucket_relock_success,
449     "Number of successful removals after relocking");
450 static long zap_bucket_fail;
451 DEBUGNODE_ULONG(zap_bucket_fail, zap_bucket_fail, "");
452 static long zap_bucket_fail2;
453 DEBUGNODE_ULONG(zap_bucket_fail2, zap_bucket_fail2, "");
454 static long cache_lock_vnodes_cel_3_failures;
455 DEBUGNODE_ULONG(vnodes_cel_3_failures, cache_lock_vnodes_cel_3_failures,
456     "Number of times 3-way vnode locking failed");
457 
458 static void cache_zap_locked(struct namecache *ncp);
459 static int vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf,
460     char **freebuf, size_t *buflen);
461 static int vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
462     char **retbuf, size_t *buflen, size_t addend);
463 static int vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf,
464     char **retbuf, size_t *buflen);
465 static int vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf,
466     char **retbuf, size_t *len, size_t addend);
467 
468 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
469 
470 static inline void
471 cache_assert_vlp_locked(struct mtx *vlp)
472 {
473 
474 	if (vlp != NULL)
475 		mtx_assert(vlp, MA_OWNED);
476 }
477 
478 static inline void
479 cache_assert_vnode_locked(struct vnode *vp)
480 {
481 	struct mtx *vlp;
482 
483 	vlp = VP2VNODELOCK(vp);
484 	cache_assert_vlp_locked(vlp);
485 }
486 
487 /*
488  * Directory vnodes with entries are held for two reasons:
489  * 1. make them less of a target for reclamation in vnlru
490  * 2. suffer smaller performance penalty in locked lookup as requeieing is avoided
491  *
492  * It will be feasible to stop doing it altogether if all filesystems start
493  * supporting lockless lookup.
494  */
495 static void
496 cache_hold_vnode(struct vnode *vp)
497 {
498 
499 	cache_assert_vnode_locked(vp);
500 	VNPASS(LIST_EMPTY(&vp->v_cache_src), vp);
501 	vhold(vp);
502 	counter_u64_add(numcachehv, 1);
503 }
504 
505 static void
506 cache_drop_vnode(struct vnode *vp)
507 {
508 
509 	/*
510 	 * Called after all locks are dropped, meaning we can't assert
511 	 * on the state of v_cache_src.
512 	 */
513 	vdrop(vp);
514 	counter_u64_add(numcachehv, -1);
515 }
516 
517 /*
518  * UMA zones.
519  */
520 static uma_zone_t __read_mostly cache_zone_small;
521 static uma_zone_t __read_mostly cache_zone_small_ts;
522 static uma_zone_t __read_mostly cache_zone_large;
523 static uma_zone_t __read_mostly cache_zone_large_ts;
524 
525 char *
526 cache_symlink_alloc(size_t size, int flags)
527 {
528 
529 	if (size < CACHE_ZONE_SMALL_SIZE) {
530 		return (uma_zalloc_smr(cache_zone_small, flags));
531 	}
532 	if (size < CACHE_ZONE_LARGE_SIZE) {
533 		return (uma_zalloc_smr(cache_zone_large, flags));
534 	}
535 	counter_u64_add(symlinktoobig, 1);
536 	SDT_PROBE1(vfs, namecache, symlink, alloc__fail, size);
537 	return (NULL);
538 }
539 
540 void
541 cache_symlink_free(char *string, size_t size)
542 {
543 
544 	MPASS(string != NULL);
545 	KASSERT(size < CACHE_ZONE_LARGE_SIZE,
546 	    ("%s: size %zu too big", __func__, size));
547 
548 	if (size < CACHE_ZONE_SMALL_SIZE) {
549 		uma_zfree_smr(cache_zone_small, string);
550 		return;
551 	}
552 	if (size < CACHE_ZONE_LARGE_SIZE) {
553 		uma_zfree_smr(cache_zone_large, string);
554 		return;
555 	}
556 	__assert_unreachable();
557 }
558 
559 static struct namecache *
560 cache_alloc_uma(int len, bool ts)
561 {
562 	struct namecache_ts *ncp_ts;
563 	struct namecache *ncp;
564 
565 	if (__predict_false(ts)) {
566 		if (len <= CACHE_PATH_CUTOFF)
567 			ncp_ts = uma_zalloc_smr(cache_zone_small_ts, M_WAITOK);
568 		else
569 			ncp_ts = uma_zalloc_smr(cache_zone_large_ts, M_WAITOK);
570 		ncp = &ncp_ts->nc_nc;
571 	} else {
572 		if (len <= CACHE_PATH_CUTOFF)
573 			ncp = uma_zalloc_smr(cache_zone_small, M_WAITOK);
574 		else
575 			ncp = uma_zalloc_smr(cache_zone_large, M_WAITOK);
576 	}
577 	return (ncp);
578 }
579 
580 static void
581 cache_free_uma(struct namecache *ncp)
582 {
583 	struct namecache_ts *ncp_ts;
584 
585 	if (__predict_false(ncp->nc_flag & NCF_TS)) {
586 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
587 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
588 			uma_zfree_smr(cache_zone_small_ts, ncp_ts);
589 		else
590 			uma_zfree_smr(cache_zone_large_ts, ncp_ts);
591 	} else {
592 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
593 			uma_zfree_smr(cache_zone_small, ncp);
594 		else
595 			uma_zfree_smr(cache_zone_large, ncp);
596 	}
597 }
598 
599 static struct namecache *
600 cache_alloc(int len, bool ts)
601 {
602 	u_long lnumcache;
603 
604 	/*
605 	 * Avoid blowout in namecache entries.
606 	 *
607 	 * Bugs:
608 	 * 1. filesystems may end up trying to add an already existing entry
609 	 * (for example this can happen after a cache miss during concurrent
610 	 * lookup), in which case we will call cache_neg_evict despite not
611 	 * adding anything.
612 	 * 2. the routine may fail to free anything and no provisions are made
613 	 * to make it try harder (see the inside for failure modes)
614 	 * 3. it only ever looks at negative entries.
615 	 */
616 	lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
617 	if (cache_neg_evict_cond(lnumcache)) {
618 		lnumcache = atomic_load_long(&numcache);
619 	}
620 	if (__predict_false(lnumcache >= ncsize)) {
621 		atomic_subtract_long(&numcache, 1);
622 		counter_u64_add(numdrops, 1);
623 		return (NULL);
624 	}
625 	return (cache_alloc_uma(len, ts));
626 }
627 
628 static void
629 cache_free(struct namecache *ncp)
630 {
631 
632 	MPASS(ncp != NULL);
633 	if ((ncp->nc_flag & NCF_DVDROP) != 0) {
634 		cache_drop_vnode(ncp->nc_dvp);
635 	}
636 	cache_free_uma(ncp);
637 	atomic_subtract_long(&numcache, 1);
638 }
639 
640 static void
641 cache_free_batch(struct cache_freebatch *batch)
642 {
643 	struct namecache *ncp, *nnp;
644 	int i;
645 
646 	i = 0;
647 	if (TAILQ_EMPTY(batch))
648 		goto out;
649 	TAILQ_FOREACH_SAFE(ncp, batch, nc_dst, nnp) {
650 		if ((ncp->nc_flag & NCF_DVDROP) != 0) {
651 			cache_drop_vnode(ncp->nc_dvp);
652 		}
653 		cache_free_uma(ncp);
654 		i++;
655 	}
656 	atomic_subtract_long(&numcache, i);
657 out:
658 	SDT_PROBE1(vfs, namecache, purge, batch, i);
659 }
660 
661 /*
662  * Hashing.
663  *
664  * The code was made to use FNV in 2001 and this choice needs to be revisited.
665  *
666  * Short summary of the difficulty:
667  * The longest name which can be inserted is NAME_MAX characters in length (or
668  * 255 at the time of writing this comment), while majority of names used in
669  * practice are significantly shorter (mostly below 10). More importantly
670  * majority of lookups performed find names are even shorter than that.
671  *
672  * This poses a problem where hashes which do better than FNV past word size
673  * (or so) tend to come with additional overhead when finalizing the result,
674  * making them noticeably slower for the most commonly used range.
675  *
676  * Consider a path like: /usr/obj/usr/src/sys/amd64/GENERIC/vnode_if.c
677  *
678  * When looking it up the most time consuming part by a large margin (at least
679  * on amd64) is hashing.  Replacing FNV with something which pessimizes short
680  * input would make the slowest part stand out even more.
681  */
682 
683 /*
684  * TODO: With the value stored we can do better than computing the hash based
685  * on the address.
686  */
687 static void
688 cache_prehash(struct vnode *vp)
689 {
690 
691 	vp->v_nchash = fnv_32_buf(&vp, sizeof(vp), FNV1_32_INIT);
692 }
693 
694 static uint32_t
695 cache_get_hash(char *name, u_char len, struct vnode *dvp)
696 {
697 
698 	return (fnv_32_buf(name, len, dvp->v_nchash));
699 }
700 
701 static uint32_t
702 cache_get_hash_iter_start(struct vnode *dvp)
703 {
704 
705 	return (dvp->v_nchash);
706 }
707 
708 static uint32_t
709 cache_get_hash_iter(char c, uint32_t hash)
710 {
711 
712 	return (fnv_32_buf(&c, 1, hash));
713 }
714 
715 static uint32_t
716 cache_get_hash_iter_finish(uint32_t hash)
717 {
718 
719 	return (hash);
720 }
721 
722 static inline struct nchashhead *
723 NCP2BUCKET(struct namecache *ncp)
724 {
725 	uint32_t hash;
726 
727 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
728 	return (NCHHASH(hash));
729 }
730 
731 static inline struct mtx *
732 NCP2BUCKETLOCK(struct namecache *ncp)
733 {
734 	uint32_t hash;
735 
736 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
737 	return (HASH2BUCKETLOCK(hash));
738 }
739 
740 #ifdef INVARIANTS
741 static void
742 cache_assert_bucket_locked(struct namecache *ncp)
743 {
744 	struct mtx *blp;
745 
746 	blp = NCP2BUCKETLOCK(ncp);
747 	mtx_assert(blp, MA_OWNED);
748 }
749 
750 static void
751 cache_assert_bucket_unlocked(struct namecache *ncp)
752 {
753 	struct mtx *blp;
754 
755 	blp = NCP2BUCKETLOCK(ncp);
756 	mtx_assert(blp, MA_NOTOWNED);
757 }
758 #else
759 #define cache_assert_bucket_locked(x) do { } while (0)
760 #define cache_assert_bucket_unlocked(x) do { } while (0)
761 #endif
762 
763 #define cache_sort_vnodes(x, y)	_cache_sort_vnodes((void **)(x), (void **)(y))
764 static void
765 _cache_sort_vnodes(void **p1, void **p2)
766 {
767 	void *tmp;
768 
769 	MPASS(*p1 != NULL || *p2 != NULL);
770 
771 	if (*p1 > *p2) {
772 		tmp = *p2;
773 		*p2 = *p1;
774 		*p1 = tmp;
775 	}
776 }
777 
778 static void
779 cache_lock_all_buckets(void)
780 {
781 	u_int i;
782 
783 	for (i = 0; i < numbucketlocks; i++)
784 		mtx_lock(&bucketlocks[i]);
785 }
786 
787 static void
788 cache_unlock_all_buckets(void)
789 {
790 	u_int i;
791 
792 	for (i = 0; i < numbucketlocks; i++)
793 		mtx_unlock(&bucketlocks[i]);
794 }
795 
796 static void
797 cache_lock_all_vnodes(void)
798 {
799 	u_int i;
800 
801 	for (i = 0; i < numvnodelocks; i++)
802 		mtx_lock(&vnodelocks[i]);
803 }
804 
805 static void
806 cache_unlock_all_vnodes(void)
807 {
808 	u_int i;
809 
810 	for (i = 0; i < numvnodelocks; i++)
811 		mtx_unlock(&vnodelocks[i]);
812 }
813 
814 static int
815 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
816 {
817 
818 	cache_sort_vnodes(&vlp1, &vlp2);
819 
820 	if (vlp1 != NULL) {
821 		if (!mtx_trylock(vlp1))
822 			return (EAGAIN);
823 	}
824 	if (!mtx_trylock(vlp2)) {
825 		if (vlp1 != NULL)
826 			mtx_unlock(vlp1);
827 		return (EAGAIN);
828 	}
829 
830 	return (0);
831 }
832 
833 static void
834 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
835 {
836 
837 	MPASS(vlp1 != NULL || vlp2 != NULL);
838 	MPASS(vlp1 <= vlp2);
839 
840 	if (vlp1 != NULL)
841 		mtx_lock(vlp1);
842 	if (vlp2 != NULL)
843 		mtx_lock(vlp2);
844 }
845 
846 static void
847 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
848 {
849 
850 	MPASS(vlp1 != NULL || vlp2 != NULL);
851 
852 	if (vlp1 != NULL)
853 		mtx_unlock(vlp1);
854 	if (vlp2 != NULL)
855 		mtx_unlock(vlp2);
856 }
857 
858 static int
859 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
860 {
861 	struct nchstats snap;
862 
863 	if (req->oldptr == NULL)
864 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
865 
866 	snap = nchstats;
867 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
868 	snap.ncs_neghits = counter_u64_fetch(numneghits);
869 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
870 	    counter_u64_fetch(numnegzaps);
871 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
872 	    counter_u64_fetch(nummiss);
873 
874 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
875 }
876 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
877     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
878     "VFS cache effectiveness statistics");
879 
880 static void
881 cache_recalc_neg_min(u_int val)
882 {
883 
884 	neg_min = (ncsize * val) / 100;
885 }
886 
887 static int
888 sysctl_negminpct(SYSCTL_HANDLER_ARGS)
889 {
890 	u_int val;
891 	int error;
892 
893 	val = ncnegminpct;
894 	error = sysctl_handle_int(oidp, &val, 0, req);
895 	if (error != 0 || req->newptr == NULL)
896 		return (error);
897 
898 	if (val == ncnegminpct)
899 		return (0);
900 	if (val < 0 || val > 99)
901 		return (EINVAL);
902 	ncnegminpct = val;
903 	cache_recalc_neg_min(val);
904 	return (0);
905 }
906 
907 SYSCTL_PROC(_vfs_cache_param, OID_AUTO, negminpct,
908     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_negminpct,
909     "I", "Negative entry \% of namecache capacity above which automatic eviction is allowed");
910 
911 #ifdef DIAGNOSTIC
912 /*
913  * Grab an atomic snapshot of the name cache hash chain lengths
914  */
915 static SYSCTL_NODE(_debug, OID_AUTO, hashstat,
916     CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
917     "hash table stats");
918 
919 static int
920 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
921 {
922 	struct nchashhead *ncpp;
923 	struct namecache *ncp;
924 	int i, error, n_nchash, *cntbuf;
925 
926 retry:
927 	n_nchash = nchash + 1;	/* nchash is max index, not count */
928 	if (req->oldptr == NULL)
929 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
930 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
931 	cache_lock_all_buckets();
932 	if (n_nchash != nchash + 1) {
933 		cache_unlock_all_buckets();
934 		free(cntbuf, M_TEMP);
935 		goto retry;
936 	}
937 	/* Scan hash tables counting entries */
938 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
939 		CK_SLIST_FOREACH(ncp, ncpp, nc_hash)
940 			cntbuf[i]++;
941 	cache_unlock_all_buckets();
942 	for (error = 0, i = 0; i < n_nchash; i++)
943 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
944 			break;
945 	free(cntbuf, M_TEMP);
946 	return (error);
947 }
948 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
949     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
950     "nchash chain lengths");
951 
952 static int
953 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
954 {
955 	int error;
956 	struct nchashhead *ncpp;
957 	struct namecache *ncp;
958 	int n_nchash;
959 	int count, maxlength, used, pct;
960 
961 	if (!req->oldptr)
962 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
963 
964 	cache_lock_all_buckets();
965 	n_nchash = nchash + 1;	/* nchash is max index, not count */
966 	used = 0;
967 	maxlength = 0;
968 
969 	/* Scan hash tables for applicable entries */
970 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
971 		count = 0;
972 		CK_SLIST_FOREACH(ncp, ncpp, nc_hash) {
973 			count++;
974 		}
975 		if (count)
976 			used++;
977 		if (maxlength < count)
978 			maxlength = count;
979 	}
980 	n_nchash = nchash + 1;
981 	cache_unlock_all_buckets();
982 	pct = (used * 100) / (n_nchash / 100);
983 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
984 	if (error)
985 		return (error);
986 	error = SYSCTL_OUT(req, &used, sizeof(used));
987 	if (error)
988 		return (error);
989 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
990 	if (error)
991 		return (error);
992 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
993 	if (error)
994 		return (error);
995 	return (0);
996 }
997 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
998     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
999     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
1000 #endif
1001 
1002 /*
1003  * Negative entries management
1004  *
1005  * Various workloads create plenty of negative entries and barely use them
1006  * afterwards. Moreover malicious users can keep performing bogus lookups
1007  * adding even more entries. For example "make tinderbox" as of writing this
1008  * comment ends up with 2.6M namecache entries in total, 1.2M of which are
1009  * negative.
1010  *
1011  * As such, a rather aggressive eviction method is needed. The currently
1012  * employed method is a placeholder.
1013  *
1014  * Entries are split over numneglists separate lists, each of which is further
1015  * split into hot and cold entries. Entries get promoted after getting a hit.
1016  * Eviction happens on addition of new entry.
1017  */
1018 static SYSCTL_NODE(_vfs_cache, OID_AUTO, neg, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1019     "Name cache negative entry statistics");
1020 
1021 SYSCTL_ULONG(_vfs_cache_neg, OID_AUTO, count, CTLFLAG_RD, &numneg, 0,
1022     "Number of negative cache entries");
1023 
1024 static COUNTER_U64_DEFINE_EARLY(neg_created);
1025 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, created, CTLFLAG_RD, &neg_created,
1026     "Number of created negative entries");
1027 
1028 static COUNTER_U64_DEFINE_EARLY(neg_evicted);
1029 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evicted, CTLFLAG_RD, &neg_evicted,
1030     "Number of evicted negative entries");
1031 
1032 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_empty);
1033 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_empty, CTLFLAG_RD,
1034     &neg_evict_skipped_empty,
1035     "Number of times evicting failed due to lack of entries");
1036 
1037 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_missed);
1038 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_missed, CTLFLAG_RD,
1039     &neg_evict_skipped_missed,
1040     "Number of times evicting failed due to target entry disappearing");
1041 
1042 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_contended);
1043 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_contended, CTLFLAG_RD,
1044     &neg_evict_skipped_contended,
1045     "Number of times evicting failed due to contention");
1046 
1047 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, hits, CTLFLAG_RD, &numneghits,
1048     "Number of cache hits (negative)");
1049 
1050 static int
1051 sysctl_neg_hot(SYSCTL_HANDLER_ARGS)
1052 {
1053 	int i, out;
1054 
1055 	out = 0;
1056 	for (i = 0; i < numneglists; i++)
1057 		out += neglists[i].nl_hotnum;
1058 
1059 	return (SYSCTL_OUT(req, &out, sizeof(out)));
1060 }
1061 SYSCTL_PROC(_vfs_cache_neg, OID_AUTO, hot, CTLTYPE_INT | CTLFLAG_RD |
1062     CTLFLAG_MPSAFE, 0, 0, sysctl_neg_hot, "I",
1063     "Number of hot negative entries");
1064 
1065 static void
1066 cache_neg_init(struct namecache *ncp)
1067 {
1068 	struct negstate *ns;
1069 
1070 	ncp->nc_flag |= NCF_NEGATIVE;
1071 	ns = NCP2NEGSTATE(ncp);
1072 	ns->neg_flag = 0;
1073 	ns->neg_hit = 0;
1074 	counter_u64_add(neg_created, 1);
1075 }
1076 
1077 #define CACHE_NEG_PROMOTION_THRESH 2
1078 
1079 static bool
1080 cache_neg_hit_prep(struct namecache *ncp)
1081 {
1082 	struct negstate *ns;
1083 	u_char n;
1084 
1085 	ns = NCP2NEGSTATE(ncp);
1086 	n = atomic_load_char(&ns->neg_hit);
1087 	for (;;) {
1088 		if (n >= CACHE_NEG_PROMOTION_THRESH)
1089 			return (false);
1090 		if (atomic_fcmpset_8(&ns->neg_hit, &n, n + 1))
1091 			break;
1092 	}
1093 	return (n + 1 == CACHE_NEG_PROMOTION_THRESH);
1094 }
1095 
1096 /*
1097  * Nothing to do here but it is provided for completeness as some
1098  * cache_neg_hit_prep callers may end up returning without even
1099  * trying to promote.
1100  */
1101 #define cache_neg_hit_abort(ncp)	do { } while (0)
1102 
1103 static void
1104 cache_neg_hit_finish(struct namecache *ncp)
1105 {
1106 
1107 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, ncp->nc_dvp, ncp->nc_name);
1108 	counter_u64_add(numneghits, 1);
1109 }
1110 
1111 /*
1112  * Move a negative entry to the hot list.
1113  */
1114 static void
1115 cache_neg_promote_locked(struct namecache *ncp)
1116 {
1117 	struct neglist *nl;
1118 	struct negstate *ns;
1119 
1120 	ns = NCP2NEGSTATE(ncp);
1121 	nl = NCP2NEGLIST(ncp);
1122 	mtx_assert(&nl->nl_lock, MA_OWNED);
1123 	if ((ns->neg_flag & NEG_HOT) == 0) {
1124 		TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1125 		TAILQ_INSERT_TAIL(&nl->nl_hotlist, ncp, nc_dst);
1126 		nl->nl_hotnum++;
1127 		ns->neg_flag |= NEG_HOT;
1128 	}
1129 }
1130 
1131 /*
1132  * Move a hot negative entry to the cold list.
1133  */
1134 static void
1135 cache_neg_demote_locked(struct namecache *ncp)
1136 {
1137 	struct neglist *nl;
1138 	struct negstate *ns;
1139 
1140 	ns = NCP2NEGSTATE(ncp);
1141 	nl = NCP2NEGLIST(ncp);
1142 	mtx_assert(&nl->nl_lock, MA_OWNED);
1143 	MPASS(ns->neg_flag & NEG_HOT);
1144 	TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1145 	TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1146 	nl->nl_hotnum--;
1147 	ns->neg_flag &= ~NEG_HOT;
1148 	atomic_store_char(&ns->neg_hit, 0);
1149 }
1150 
1151 /*
1152  * Move a negative entry to the hot list if it matches the lookup.
1153  *
1154  * We have to take locks, but they may be contended and in the worst
1155  * case we may need to go off CPU. We don't want to spin within the
1156  * smr section and we can't block with it. Exiting the section means
1157  * the found entry could have been evicted. We are going to look it
1158  * up again.
1159  */
1160 static bool
1161 cache_neg_promote_cond(struct vnode *dvp, struct componentname *cnp,
1162     struct namecache *oncp, uint32_t hash)
1163 {
1164 	struct namecache *ncp;
1165 	struct neglist *nl;
1166 	u_char nc_flag;
1167 
1168 	nl = NCP2NEGLIST(oncp);
1169 
1170 	mtx_lock(&nl->nl_lock);
1171 	/*
1172 	 * For hash iteration.
1173 	 */
1174 	vfs_smr_enter();
1175 
1176 	/*
1177 	 * Avoid all surprises by only succeeding if we got the same entry and
1178 	 * bailing completely otherwise.
1179 	 * XXX There are no provisions to keep the vnode around, meaning we may
1180 	 * end up promoting a negative entry for a *new* vnode and returning
1181 	 * ENOENT on its account. This is the error we want to return anyway
1182 	 * and promotion is harmless.
1183 	 *
1184 	 * In particular at this point there can be a new ncp which matches the
1185 	 * search but hashes to a different neglist.
1186 	 */
1187 	CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1188 		if (ncp == oncp)
1189 			break;
1190 	}
1191 
1192 	/*
1193 	 * No match to begin with.
1194 	 */
1195 	if (__predict_false(ncp == NULL)) {
1196 		goto out_abort;
1197 	}
1198 
1199 	/*
1200 	 * The newly found entry may be something different...
1201 	 */
1202 	if (!(ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1203 	    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))) {
1204 		goto out_abort;
1205 	}
1206 
1207 	/*
1208 	 * ... and not even negative.
1209 	 */
1210 	nc_flag = atomic_load_char(&ncp->nc_flag);
1211 	if ((nc_flag & NCF_NEGATIVE) == 0) {
1212 		goto out_abort;
1213 	}
1214 
1215 	if (!cache_ncp_canuse(ncp)) {
1216 		goto out_abort;
1217 	}
1218 
1219 	cache_neg_promote_locked(ncp);
1220 	cache_neg_hit_finish(ncp);
1221 	vfs_smr_exit();
1222 	mtx_unlock(&nl->nl_lock);
1223 	return (true);
1224 out_abort:
1225 	vfs_smr_exit();
1226 	mtx_unlock(&nl->nl_lock);
1227 	return (false);
1228 }
1229 
1230 static void
1231 cache_neg_promote(struct namecache *ncp)
1232 {
1233 	struct neglist *nl;
1234 
1235 	nl = NCP2NEGLIST(ncp);
1236 	mtx_lock(&nl->nl_lock);
1237 	cache_neg_promote_locked(ncp);
1238 	mtx_unlock(&nl->nl_lock);
1239 }
1240 
1241 static void
1242 cache_neg_insert(struct namecache *ncp)
1243 {
1244 	struct neglist *nl;
1245 
1246 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
1247 	cache_assert_bucket_locked(ncp);
1248 	nl = NCP2NEGLIST(ncp);
1249 	mtx_lock(&nl->nl_lock);
1250 	TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1251 	mtx_unlock(&nl->nl_lock);
1252 	atomic_add_long(&numneg, 1);
1253 }
1254 
1255 static void
1256 cache_neg_remove(struct namecache *ncp)
1257 {
1258 	struct neglist *nl;
1259 	struct negstate *ns;
1260 
1261 	cache_assert_bucket_locked(ncp);
1262 	nl = NCP2NEGLIST(ncp);
1263 	ns = NCP2NEGSTATE(ncp);
1264 	mtx_lock(&nl->nl_lock);
1265 	if ((ns->neg_flag & NEG_HOT) != 0) {
1266 		TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1267 		nl->nl_hotnum--;
1268 	} else {
1269 		TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1270 	}
1271 	mtx_unlock(&nl->nl_lock);
1272 	atomic_subtract_long(&numneg, 1);
1273 }
1274 
1275 static struct neglist *
1276 cache_neg_evict_select_list(void)
1277 {
1278 	struct neglist *nl;
1279 	u_int c;
1280 
1281 	c = atomic_fetchadd_int(&neg_cycle, 1) + 1;
1282 	nl = &neglists[c % numneglists];
1283 	if (!mtx_trylock(&nl->nl_evict_lock)) {
1284 		counter_u64_add(neg_evict_skipped_contended, 1);
1285 		return (NULL);
1286 	}
1287 	return (nl);
1288 }
1289 
1290 static struct namecache *
1291 cache_neg_evict_select_entry(struct neglist *nl)
1292 {
1293 	struct namecache *ncp, *lncp;
1294 	struct negstate *ns, *lns;
1295 	int i;
1296 
1297 	mtx_assert(&nl->nl_evict_lock, MA_OWNED);
1298 	mtx_assert(&nl->nl_lock, MA_OWNED);
1299 	ncp = TAILQ_FIRST(&nl->nl_list);
1300 	if (ncp == NULL)
1301 		return (NULL);
1302 	lncp = ncp;
1303 	lns = NCP2NEGSTATE(lncp);
1304 	for (i = 1; i < 4; i++) {
1305 		ncp = TAILQ_NEXT(ncp, nc_dst);
1306 		if (ncp == NULL)
1307 			break;
1308 		ns = NCP2NEGSTATE(ncp);
1309 		if (ns->neg_hit < lns->neg_hit) {
1310 			lncp = ncp;
1311 			lns = ns;
1312 		}
1313 	}
1314 	return (lncp);
1315 }
1316 
1317 static bool
1318 cache_neg_evict(void)
1319 {
1320 	struct namecache *ncp, *ncp2;
1321 	struct neglist *nl;
1322 	struct vnode *dvp;
1323 	struct mtx *dvlp;
1324 	struct mtx *blp;
1325 	uint32_t hash;
1326 	u_char nlen;
1327 	bool evicted;
1328 
1329 	nl = cache_neg_evict_select_list();
1330 	if (nl == NULL) {
1331 		return (false);
1332 	}
1333 
1334 	mtx_lock(&nl->nl_lock);
1335 	ncp = TAILQ_FIRST(&nl->nl_hotlist);
1336 	if (ncp != NULL) {
1337 		cache_neg_demote_locked(ncp);
1338 	}
1339 	ncp = cache_neg_evict_select_entry(nl);
1340 	if (ncp == NULL) {
1341 		counter_u64_add(neg_evict_skipped_empty, 1);
1342 		mtx_unlock(&nl->nl_lock);
1343 		mtx_unlock(&nl->nl_evict_lock);
1344 		return (false);
1345 	}
1346 	nlen = ncp->nc_nlen;
1347 	dvp = ncp->nc_dvp;
1348 	hash = cache_get_hash(ncp->nc_name, nlen, dvp);
1349 	dvlp = VP2VNODELOCK(dvp);
1350 	blp = HASH2BUCKETLOCK(hash);
1351 	mtx_unlock(&nl->nl_lock);
1352 	mtx_unlock(&nl->nl_evict_lock);
1353 	mtx_lock(dvlp);
1354 	mtx_lock(blp);
1355 	/*
1356 	 * Note that since all locks were dropped above, the entry may be
1357 	 * gone or reallocated to be something else.
1358 	 */
1359 	CK_SLIST_FOREACH(ncp2, (NCHHASH(hash)), nc_hash) {
1360 		if (ncp2 == ncp && ncp2->nc_dvp == dvp &&
1361 		    ncp2->nc_nlen == nlen && (ncp2->nc_flag & NCF_NEGATIVE) != 0)
1362 			break;
1363 	}
1364 	if (ncp2 == NULL) {
1365 		counter_u64_add(neg_evict_skipped_missed, 1);
1366 		ncp = NULL;
1367 		evicted = false;
1368 	} else {
1369 		MPASS(dvlp == VP2VNODELOCK(ncp->nc_dvp));
1370 		MPASS(blp == NCP2BUCKETLOCK(ncp));
1371 		SDT_PROBE2(vfs, namecache, evict_negative, done, ncp->nc_dvp,
1372 		    ncp->nc_name);
1373 		cache_zap_locked(ncp);
1374 		counter_u64_add(neg_evicted, 1);
1375 		evicted = true;
1376 	}
1377 	mtx_unlock(blp);
1378 	mtx_unlock(dvlp);
1379 	if (ncp != NULL)
1380 		cache_free(ncp);
1381 	return (evicted);
1382 }
1383 
1384 /*
1385  * Maybe evict a negative entry to create more room.
1386  *
1387  * The ncnegfactor parameter limits what fraction of the total count
1388  * can comprise of negative entries. However, if the cache is just
1389  * warming up this leads to excessive evictions.  As such, ncnegminpct
1390  * (recomputed to neg_min) dictates whether the above should be
1391  * applied.
1392  *
1393  * Try evicting if the cache is close to full capacity regardless of
1394  * other considerations.
1395  */
1396 static bool
1397 cache_neg_evict_cond(u_long lnumcache)
1398 {
1399 	u_long lnumneg;
1400 
1401 	if (ncsize - 1000 < lnumcache)
1402 		goto out_evict;
1403 	lnumneg = atomic_load_long(&numneg);
1404 	if (lnumneg < neg_min)
1405 		return (false);
1406 	if (lnumneg * ncnegfactor < lnumcache)
1407 		return (false);
1408 out_evict:
1409 	return (cache_neg_evict());
1410 }
1411 
1412 /*
1413  * cache_zap_locked():
1414  *
1415  *   Removes a namecache entry from cache, whether it contains an actual
1416  *   pointer to a vnode or if it is just a negative cache entry.
1417  */
1418 static void
1419 cache_zap_locked(struct namecache *ncp)
1420 {
1421 	struct nchashhead *ncpp;
1422 	struct vnode *dvp, *vp;
1423 
1424 	dvp = ncp->nc_dvp;
1425 	vp = ncp->nc_vp;
1426 
1427 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1428 		cache_assert_vnode_locked(vp);
1429 	cache_assert_vnode_locked(dvp);
1430 	cache_assert_bucket_locked(ncp);
1431 
1432 	cache_ncp_invalidate(ncp);
1433 
1434 	ncpp = NCP2BUCKET(ncp);
1435 	CK_SLIST_REMOVE(ncpp, ncp, namecache, nc_hash);
1436 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1437 		SDT_PROBE3(vfs, namecache, zap, done, dvp, ncp->nc_name, vp);
1438 		TAILQ_REMOVE(&vp->v_cache_dst, ncp, nc_dst);
1439 		if (ncp == vp->v_cache_dd) {
1440 			atomic_store_ptr(&vp->v_cache_dd, NULL);
1441 		}
1442 	} else {
1443 		SDT_PROBE2(vfs, namecache, zap_negative, done, dvp, ncp->nc_name);
1444 		cache_neg_remove(ncp);
1445 	}
1446 	if (ncp->nc_flag & NCF_ISDOTDOT) {
1447 		if (ncp == dvp->v_cache_dd) {
1448 			atomic_store_ptr(&dvp->v_cache_dd, NULL);
1449 		}
1450 	} else {
1451 		LIST_REMOVE(ncp, nc_src);
1452 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1453 			ncp->nc_flag |= NCF_DVDROP;
1454 		}
1455 	}
1456 }
1457 
1458 static void
1459 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
1460 {
1461 	struct mtx *blp;
1462 
1463 	MPASS(ncp->nc_dvp == vp);
1464 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
1465 	cache_assert_vnode_locked(vp);
1466 
1467 	blp = NCP2BUCKETLOCK(ncp);
1468 	mtx_lock(blp);
1469 	cache_zap_locked(ncp);
1470 	mtx_unlock(blp);
1471 }
1472 
1473 static bool
1474 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
1475     struct mtx **vlpp)
1476 {
1477 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
1478 	struct mtx *blp;
1479 
1480 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
1481 	cache_assert_vnode_locked(vp);
1482 
1483 	if (ncp->nc_flag & NCF_NEGATIVE) {
1484 		if (*vlpp != NULL) {
1485 			mtx_unlock(*vlpp);
1486 			*vlpp = NULL;
1487 		}
1488 		cache_zap_negative_locked_vnode_kl(ncp, vp);
1489 		return (true);
1490 	}
1491 
1492 	pvlp = VP2VNODELOCK(vp);
1493 	blp = NCP2BUCKETLOCK(ncp);
1494 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
1495 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
1496 
1497 	if (*vlpp == vlp1 || *vlpp == vlp2) {
1498 		to_unlock = *vlpp;
1499 		*vlpp = NULL;
1500 	} else {
1501 		if (*vlpp != NULL) {
1502 			mtx_unlock(*vlpp);
1503 			*vlpp = NULL;
1504 		}
1505 		cache_sort_vnodes(&vlp1, &vlp2);
1506 		if (vlp1 == pvlp) {
1507 			mtx_lock(vlp2);
1508 			to_unlock = vlp2;
1509 		} else {
1510 			if (!mtx_trylock(vlp1))
1511 				goto out_relock;
1512 			to_unlock = vlp1;
1513 		}
1514 	}
1515 	mtx_lock(blp);
1516 	cache_zap_locked(ncp);
1517 	mtx_unlock(blp);
1518 	if (to_unlock != NULL)
1519 		mtx_unlock(to_unlock);
1520 	return (true);
1521 
1522 out_relock:
1523 	mtx_unlock(vlp2);
1524 	mtx_lock(vlp1);
1525 	mtx_lock(vlp2);
1526 	MPASS(*vlpp == NULL);
1527 	*vlpp = vlp1;
1528 	return (false);
1529 }
1530 
1531 /*
1532  * If trylocking failed we can get here. We know enough to take all needed locks
1533  * in the right order and re-lookup the entry.
1534  */
1535 static int
1536 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1537     struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1538     struct mtx *blp)
1539 {
1540 	struct namecache *rncp;
1541 
1542 	cache_assert_bucket_unlocked(ncp);
1543 
1544 	cache_sort_vnodes(&dvlp, &vlp);
1545 	cache_lock_vnodes(dvlp, vlp);
1546 	mtx_lock(blp);
1547 	CK_SLIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1548 		if (rncp == ncp && rncp->nc_dvp == dvp &&
1549 		    rncp->nc_nlen == cnp->cn_namelen &&
1550 		    !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1551 			break;
1552 	}
1553 	if (rncp != NULL) {
1554 		cache_zap_locked(rncp);
1555 		mtx_unlock(blp);
1556 		cache_unlock_vnodes(dvlp, vlp);
1557 		counter_u64_add(zap_bucket_relock_success, 1);
1558 		return (0);
1559 	}
1560 
1561 	mtx_unlock(blp);
1562 	cache_unlock_vnodes(dvlp, vlp);
1563 	return (EAGAIN);
1564 }
1565 
1566 static int __noinline
1567 cache_zap_locked_bucket(struct namecache *ncp, struct componentname *cnp,
1568     uint32_t hash, struct mtx *blp)
1569 {
1570 	struct mtx *dvlp, *vlp;
1571 	struct vnode *dvp;
1572 
1573 	cache_assert_bucket_locked(ncp);
1574 
1575 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1576 	vlp = NULL;
1577 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1578 		vlp = VP2VNODELOCK(ncp->nc_vp);
1579 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1580 		cache_zap_locked(ncp);
1581 		mtx_unlock(blp);
1582 		cache_unlock_vnodes(dvlp, vlp);
1583 		return (0);
1584 	}
1585 
1586 	dvp = ncp->nc_dvp;
1587 	mtx_unlock(blp);
1588 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1589 }
1590 
1591 static __noinline int
1592 cache_remove_cnp(struct vnode *dvp, struct componentname *cnp)
1593 {
1594 	struct namecache *ncp;
1595 	struct mtx *blp;
1596 	struct mtx *dvlp, *dvlp2;
1597 	uint32_t hash;
1598 	int error;
1599 
1600 	if (cnp->cn_namelen == 2 &&
1601 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1602 		dvlp = VP2VNODELOCK(dvp);
1603 		dvlp2 = NULL;
1604 		mtx_lock(dvlp);
1605 retry_dotdot:
1606 		ncp = dvp->v_cache_dd;
1607 		if (ncp == NULL) {
1608 			mtx_unlock(dvlp);
1609 			if (dvlp2 != NULL)
1610 				mtx_unlock(dvlp2);
1611 			SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1612 			return (0);
1613 		}
1614 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1615 			if (!cache_zap_locked_vnode_kl2(ncp, dvp, &dvlp2))
1616 				goto retry_dotdot;
1617 			MPASS(dvp->v_cache_dd == NULL);
1618 			mtx_unlock(dvlp);
1619 			if (dvlp2 != NULL)
1620 				mtx_unlock(dvlp2);
1621 			cache_free(ncp);
1622 		} else {
1623 			atomic_store_ptr(&dvp->v_cache_dd, NULL);
1624 			mtx_unlock(dvlp);
1625 			if (dvlp2 != NULL)
1626 				mtx_unlock(dvlp2);
1627 		}
1628 		SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1629 		return (1);
1630 	}
1631 
1632 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1633 	blp = HASH2BUCKETLOCK(hash);
1634 retry:
1635 	if (CK_SLIST_EMPTY(NCHHASH(hash)))
1636 		goto out_no_entry;
1637 
1638 	mtx_lock(blp);
1639 
1640 	CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1641 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1642 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1643 			break;
1644 	}
1645 
1646 	if (ncp == NULL) {
1647 		mtx_unlock(blp);
1648 		goto out_no_entry;
1649 	}
1650 
1651 	error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
1652 	if (__predict_false(error != 0)) {
1653 		zap_bucket_fail++;
1654 		goto retry;
1655 	}
1656 	counter_u64_add(numposzaps, 1);
1657 	SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1658 	cache_free(ncp);
1659 	return (1);
1660 out_no_entry:
1661 	counter_u64_add(nummisszap, 1);
1662 	SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1663 	return (0);
1664 }
1665 
1666 static int __noinline
1667 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1668     struct timespec *tsp, int *ticksp)
1669 {
1670 	int ltype;
1671 
1672 	*vpp = dvp;
1673 	counter_u64_add(dothits, 1);
1674 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1675 	if (tsp != NULL)
1676 		timespecclear(tsp);
1677 	if (ticksp != NULL)
1678 		*ticksp = ticks;
1679 	vrefact(*vpp);
1680 	/*
1681 	 * When we lookup "." we still can be asked to lock it
1682 	 * differently...
1683 	 */
1684 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1685 	if (ltype != VOP_ISLOCKED(*vpp)) {
1686 		if (ltype == LK_EXCLUSIVE) {
1687 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1688 			if (VN_IS_DOOMED((*vpp))) {
1689 				/* forced unmount */
1690 				vrele(*vpp);
1691 				*vpp = NULL;
1692 				return (ENOENT);
1693 			}
1694 		} else
1695 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1696 	}
1697 	return (-1);
1698 }
1699 
1700 static int __noinline
1701 cache_lookup_dotdot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1702     struct timespec *tsp, int *ticksp)
1703 {
1704 	struct namecache_ts *ncp_ts;
1705 	struct namecache *ncp;
1706 	struct mtx *dvlp;
1707 	enum vgetstate vs;
1708 	int error, ltype;
1709 	bool whiteout;
1710 
1711 	MPASS((cnp->cn_flags & ISDOTDOT) != 0);
1712 
1713 	if ((cnp->cn_flags & MAKEENTRY) == 0) {
1714 		cache_remove_cnp(dvp, cnp);
1715 		return (0);
1716 	}
1717 
1718 	counter_u64_add(dotdothits, 1);
1719 retry:
1720 	dvlp = VP2VNODELOCK(dvp);
1721 	mtx_lock(dvlp);
1722 	ncp = dvp->v_cache_dd;
1723 	if (ncp == NULL) {
1724 		SDT_PROBE2(vfs, namecache, lookup, miss, dvp, "..");
1725 		mtx_unlock(dvlp);
1726 		return (0);
1727 	}
1728 	if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1729 		if (ncp->nc_flag & NCF_NEGATIVE)
1730 			*vpp = NULL;
1731 		else
1732 			*vpp = ncp->nc_vp;
1733 	} else
1734 		*vpp = ncp->nc_dvp;
1735 	if (*vpp == NULL)
1736 		goto negative_success;
1737 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..", *vpp);
1738 	cache_out_ts(ncp, tsp, ticksp);
1739 	if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1740 	    NCF_DTS && tsp != NULL) {
1741 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1742 		*tsp = ncp_ts->nc_dotdottime;
1743 	}
1744 
1745 	MPASS(dvp != *vpp);
1746 	ltype = VOP_ISLOCKED(dvp);
1747 	VOP_UNLOCK(dvp);
1748 	vs = vget_prep(*vpp);
1749 	mtx_unlock(dvlp);
1750 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1751 	vn_lock(dvp, ltype | LK_RETRY);
1752 	if (VN_IS_DOOMED(dvp)) {
1753 		if (error == 0)
1754 			vput(*vpp);
1755 		*vpp = NULL;
1756 		return (ENOENT);
1757 	}
1758 	if (error) {
1759 		*vpp = NULL;
1760 		goto retry;
1761 	}
1762 	return (-1);
1763 negative_success:
1764 	if (__predict_false(cnp->cn_nameiop == CREATE)) {
1765 		if (cnp->cn_flags & ISLASTCN) {
1766 			counter_u64_add(numnegzaps, 1);
1767 			cache_zap_negative_locked_vnode_kl(ncp, dvp);
1768 			mtx_unlock(dvlp);
1769 			cache_free(ncp);
1770 			return (0);
1771 		}
1772 	}
1773 
1774 	whiteout = (ncp->nc_flag & NCF_WHITE);
1775 	cache_out_ts(ncp, tsp, ticksp);
1776 	if (cache_neg_hit_prep(ncp))
1777 		cache_neg_promote(ncp);
1778 	else
1779 		cache_neg_hit_finish(ncp);
1780 	mtx_unlock(dvlp);
1781 	if (whiteout)
1782 		cnp->cn_flags |= ISWHITEOUT;
1783 	return (ENOENT);
1784 }
1785 
1786 /**
1787  * Lookup a name in the name cache
1788  *
1789  * # Arguments
1790  *
1791  * - dvp:	Parent directory in which to search.
1792  * - vpp:	Return argument.  Will contain desired vnode on cache hit.
1793  * - cnp:	Parameters of the name search.  The most interesting bits of
1794  *   		the cn_flags field have the following meanings:
1795  *   	- MAKEENTRY:	If clear, free an entry from the cache rather than look
1796  *   			it up.
1797  *   	- ISDOTDOT:	Must be set if and only if cn_nameptr == ".."
1798  * - tsp:	Return storage for cache timestamp.  On a successful (positive
1799  *   		or negative) lookup, tsp will be filled with any timespec that
1800  *   		was stored when this cache entry was created.  However, it will
1801  *   		be clear for "." entries.
1802  * - ticks:	Return storage for alternate cache timestamp.  On a successful
1803  *   		(positive or negative) lookup, it will contain the ticks value
1804  *   		that was current when the cache entry was created, unless cnp
1805  *   		was ".".
1806  *
1807  * Either both tsp and ticks have to be provided or neither of them.
1808  *
1809  * # Returns
1810  *
1811  * - -1:	A positive cache hit.  vpp will contain the desired vnode.
1812  * - ENOENT:	A negative cache hit, or dvp was recycled out from under us due
1813  *		to a forced unmount.  vpp will not be modified.  If the entry
1814  *		is a whiteout, then the ISWHITEOUT flag will be set in
1815  *		cnp->cn_flags.
1816  * - 0:		A cache miss.  vpp will not be modified.
1817  *
1818  * # Locking
1819  *
1820  * On a cache hit, vpp will be returned locked and ref'd.  If we're looking up
1821  * .., dvp is unlocked.  If we're looking up . an extra ref is taken, but the
1822  * lock is not recursively acquired.
1823  */
1824 static int __noinline
1825 cache_lookup_fallback(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1826     struct timespec *tsp, int *ticksp)
1827 {
1828 	struct namecache *ncp;
1829 	struct mtx *blp;
1830 	uint32_t hash;
1831 	enum vgetstate vs;
1832 	int error;
1833 	bool whiteout;
1834 
1835 	MPASS((cnp->cn_flags & ISDOTDOT) == 0);
1836 	MPASS((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) != 0);
1837 
1838 retry:
1839 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1840 	blp = HASH2BUCKETLOCK(hash);
1841 	mtx_lock(blp);
1842 
1843 	CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1844 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1845 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1846 			break;
1847 	}
1848 
1849 	if (__predict_false(ncp == NULL)) {
1850 		mtx_unlock(blp);
1851 		SDT_PROBE2(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr);
1852 		counter_u64_add(nummiss, 1);
1853 		return (0);
1854 	}
1855 
1856 	if (ncp->nc_flag & NCF_NEGATIVE)
1857 		goto negative_success;
1858 
1859 	counter_u64_add(numposhits, 1);
1860 	*vpp = ncp->nc_vp;
1861 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
1862 	cache_out_ts(ncp, tsp, ticksp);
1863 	MPASS(dvp != *vpp);
1864 	vs = vget_prep(*vpp);
1865 	mtx_unlock(blp);
1866 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1867 	if (error) {
1868 		*vpp = NULL;
1869 		goto retry;
1870 	}
1871 	return (-1);
1872 negative_success:
1873 	/*
1874 	 * We don't get here with regular lookup apart from corner cases.
1875 	 */
1876 	if (__predict_true(cnp->cn_nameiop == CREATE)) {
1877 		if (cnp->cn_flags & ISLASTCN) {
1878 			counter_u64_add(numnegzaps, 1);
1879 			error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
1880 			if (__predict_false(error != 0)) {
1881 				zap_bucket_fail2++;
1882 				goto retry;
1883 			}
1884 			cache_free(ncp);
1885 			return (0);
1886 		}
1887 	}
1888 
1889 	whiteout = (ncp->nc_flag & NCF_WHITE);
1890 	cache_out_ts(ncp, tsp, ticksp);
1891 	if (cache_neg_hit_prep(ncp))
1892 		cache_neg_promote(ncp);
1893 	else
1894 		cache_neg_hit_finish(ncp);
1895 	mtx_unlock(blp);
1896 	if (whiteout)
1897 		cnp->cn_flags |= ISWHITEOUT;
1898 	return (ENOENT);
1899 }
1900 
1901 int
1902 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1903     struct timespec *tsp, int *ticksp)
1904 {
1905 	struct namecache *ncp;
1906 	uint32_t hash;
1907 	enum vgetstate vs;
1908 	int error;
1909 	bool whiteout, neg_promote;
1910 	u_short nc_flag;
1911 
1912 	MPASS((tsp == NULL && ticksp == NULL) || (tsp != NULL && ticksp != NULL));
1913 
1914 #ifdef DEBUG_CACHE
1915 	if (__predict_false(!doingcache)) {
1916 		cnp->cn_flags &= ~MAKEENTRY;
1917 		return (0);
1918 	}
1919 #endif
1920 
1921 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
1922 		if (cnp->cn_namelen == 1)
1923 			return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1924 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.')
1925 			return (cache_lookup_dotdot(dvp, vpp, cnp, tsp, ticksp));
1926 	}
1927 
1928 	MPASS((cnp->cn_flags & ISDOTDOT) == 0);
1929 
1930 	if ((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) == 0) {
1931 		cache_remove_cnp(dvp, cnp);
1932 		return (0);
1933 	}
1934 
1935 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1936 	vfs_smr_enter();
1937 
1938 	CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1939 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1940 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1941 			break;
1942 	}
1943 
1944 	if (__predict_false(ncp == NULL)) {
1945 		vfs_smr_exit();
1946 		SDT_PROBE2(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr);
1947 		counter_u64_add(nummiss, 1);
1948 		return (0);
1949 	}
1950 
1951 	nc_flag = atomic_load_char(&ncp->nc_flag);
1952 	if (nc_flag & NCF_NEGATIVE)
1953 		goto negative_success;
1954 
1955 	counter_u64_add(numposhits, 1);
1956 	*vpp = ncp->nc_vp;
1957 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
1958 	cache_out_ts(ncp, tsp, ticksp);
1959 	MPASS(dvp != *vpp);
1960 	if (!cache_ncp_canuse(ncp)) {
1961 		vfs_smr_exit();
1962 		*vpp = NULL;
1963 		goto out_fallback;
1964 	}
1965 	vs = vget_prep_smr(*vpp);
1966 	vfs_smr_exit();
1967 	if (__predict_false(vs == VGET_NONE)) {
1968 		*vpp = NULL;
1969 		goto out_fallback;
1970 	}
1971 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1972 	if (error) {
1973 		*vpp = NULL;
1974 		goto out_fallback;
1975 	}
1976 	return (-1);
1977 negative_success:
1978 	if (cnp->cn_nameiop == CREATE) {
1979 		if (cnp->cn_flags & ISLASTCN) {
1980 			vfs_smr_exit();
1981 			goto out_fallback;
1982 		}
1983 	}
1984 
1985 	cache_out_ts(ncp, tsp, ticksp);
1986 	whiteout = (atomic_load_char(&ncp->nc_flag) & NCF_WHITE);
1987 	neg_promote = cache_neg_hit_prep(ncp);
1988 	if (!cache_ncp_canuse(ncp)) {
1989 		cache_neg_hit_abort(ncp);
1990 		vfs_smr_exit();
1991 		goto out_fallback;
1992 	}
1993 	if (neg_promote) {
1994 		vfs_smr_exit();
1995 		if (!cache_neg_promote_cond(dvp, cnp, ncp, hash))
1996 			goto out_fallback;
1997 	} else {
1998 		cache_neg_hit_finish(ncp);
1999 		vfs_smr_exit();
2000 	}
2001 	if (whiteout)
2002 		cnp->cn_flags |= ISWHITEOUT;
2003 	return (ENOENT);
2004 out_fallback:
2005 	return (cache_lookup_fallback(dvp, vpp, cnp, tsp, ticksp));
2006 }
2007 
2008 struct celockstate {
2009 	struct mtx *vlp[3];
2010 	struct mtx *blp[2];
2011 };
2012 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
2013 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
2014 
2015 static inline void
2016 cache_celockstate_init(struct celockstate *cel)
2017 {
2018 
2019 	bzero(cel, sizeof(*cel));
2020 }
2021 
2022 static void
2023 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
2024     struct vnode *dvp)
2025 {
2026 	struct mtx *vlp1, *vlp2;
2027 
2028 	MPASS(cel->vlp[0] == NULL);
2029 	MPASS(cel->vlp[1] == NULL);
2030 	MPASS(cel->vlp[2] == NULL);
2031 
2032 	MPASS(vp != NULL || dvp != NULL);
2033 
2034 	vlp1 = VP2VNODELOCK(vp);
2035 	vlp2 = VP2VNODELOCK(dvp);
2036 	cache_sort_vnodes(&vlp1, &vlp2);
2037 
2038 	if (vlp1 != NULL) {
2039 		mtx_lock(vlp1);
2040 		cel->vlp[0] = vlp1;
2041 	}
2042 	mtx_lock(vlp2);
2043 	cel->vlp[1] = vlp2;
2044 }
2045 
2046 static void
2047 cache_unlock_vnodes_cel(struct celockstate *cel)
2048 {
2049 
2050 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
2051 
2052 	if (cel->vlp[0] != NULL)
2053 		mtx_unlock(cel->vlp[0]);
2054 	if (cel->vlp[1] != NULL)
2055 		mtx_unlock(cel->vlp[1]);
2056 	if (cel->vlp[2] != NULL)
2057 		mtx_unlock(cel->vlp[2]);
2058 }
2059 
2060 static bool
2061 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
2062 {
2063 	struct mtx *vlp;
2064 	bool ret;
2065 
2066 	cache_assert_vlp_locked(cel->vlp[0]);
2067 	cache_assert_vlp_locked(cel->vlp[1]);
2068 	MPASS(cel->vlp[2] == NULL);
2069 
2070 	MPASS(vp != NULL);
2071 	vlp = VP2VNODELOCK(vp);
2072 
2073 	ret = true;
2074 	if (vlp >= cel->vlp[1]) {
2075 		mtx_lock(vlp);
2076 	} else {
2077 		if (mtx_trylock(vlp))
2078 			goto out;
2079 		cache_lock_vnodes_cel_3_failures++;
2080 		cache_unlock_vnodes_cel(cel);
2081 		if (vlp < cel->vlp[0]) {
2082 			mtx_lock(vlp);
2083 			mtx_lock(cel->vlp[0]);
2084 			mtx_lock(cel->vlp[1]);
2085 		} else {
2086 			if (cel->vlp[0] != NULL)
2087 				mtx_lock(cel->vlp[0]);
2088 			mtx_lock(vlp);
2089 			mtx_lock(cel->vlp[1]);
2090 		}
2091 		ret = false;
2092 	}
2093 out:
2094 	cel->vlp[2] = vlp;
2095 	return (ret);
2096 }
2097 
2098 static void
2099 cache_lock_buckets_cel(struct celockstate *cel, struct mtx *blp1,
2100     struct mtx *blp2)
2101 {
2102 
2103 	MPASS(cel->blp[0] == NULL);
2104 	MPASS(cel->blp[1] == NULL);
2105 
2106 	cache_sort_vnodes(&blp1, &blp2);
2107 
2108 	if (blp1 != NULL) {
2109 		mtx_lock(blp1);
2110 		cel->blp[0] = blp1;
2111 	}
2112 	mtx_lock(blp2);
2113 	cel->blp[1] = blp2;
2114 }
2115 
2116 static void
2117 cache_unlock_buckets_cel(struct celockstate *cel)
2118 {
2119 
2120 	if (cel->blp[0] != NULL)
2121 		mtx_unlock(cel->blp[0]);
2122 	mtx_unlock(cel->blp[1]);
2123 }
2124 
2125 /*
2126  * Lock part of the cache affected by the insertion.
2127  *
2128  * This means vnodelocks for dvp, vp and the relevant bucketlock.
2129  * However, insertion can result in removal of an old entry. In this
2130  * case we have an additional vnode and bucketlock pair to lock.
2131  *
2132  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
2133  * preserving the locking order (smaller address first).
2134  */
2135 static void
2136 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2137     uint32_t hash)
2138 {
2139 	struct namecache *ncp;
2140 	struct mtx *blps[2];
2141 	u_char nc_flag;
2142 
2143 	blps[0] = HASH2BUCKETLOCK(hash);
2144 	for (;;) {
2145 		blps[1] = NULL;
2146 		cache_lock_vnodes_cel(cel, dvp, vp);
2147 		if (vp == NULL || vp->v_type != VDIR)
2148 			break;
2149 		ncp = atomic_load_consume_ptr(&vp->v_cache_dd);
2150 		if (ncp == NULL)
2151 			break;
2152 		nc_flag = atomic_load_char(&ncp->nc_flag);
2153 		if ((nc_flag & NCF_ISDOTDOT) == 0)
2154 			break;
2155 		MPASS(ncp->nc_dvp == vp);
2156 		blps[1] = NCP2BUCKETLOCK(ncp);
2157 		if ((nc_flag & NCF_NEGATIVE) != 0)
2158 			break;
2159 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2160 			break;
2161 		/*
2162 		 * All vnodes got re-locked. Re-validate the state and if
2163 		 * nothing changed we are done. Otherwise restart.
2164 		 */
2165 		if (ncp == vp->v_cache_dd &&
2166 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2167 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
2168 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2169 			break;
2170 		cache_unlock_vnodes_cel(cel);
2171 		cel->vlp[0] = NULL;
2172 		cel->vlp[1] = NULL;
2173 		cel->vlp[2] = NULL;
2174 	}
2175 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
2176 }
2177 
2178 static void
2179 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2180     uint32_t hash)
2181 {
2182 	struct namecache *ncp;
2183 	struct mtx *blps[2];
2184 	u_char nc_flag;
2185 
2186 	blps[0] = HASH2BUCKETLOCK(hash);
2187 	for (;;) {
2188 		blps[1] = NULL;
2189 		cache_lock_vnodes_cel(cel, dvp, vp);
2190 		ncp = atomic_load_consume_ptr(&dvp->v_cache_dd);
2191 		if (ncp == NULL)
2192 			break;
2193 		nc_flag = atomic_load_char(&ncp->nc_flag);
2194 		if ((nc_flag & NCF_ISDOTDOT) == 0)
2195 			break;
2196 		MPASS(ncp->nc_dvp == dvp);
2197 		blps[1] = NCP2BUCKETLOCK(ncp);
2198 		if ((nc_flag & NCF_NEGATIVE) != 0)
2199 			break;
2200 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2201 			break;
2202 		if (ncp == dvp->v_cache_dd &&
2203 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2204 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
2205 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2206 			break;
2207 		cache_unlock_vnodes_cel(cel);
2208 		cel->vlp[0] = NULL;
2209 		cel->vlp[1] = NULL;
2210 		cel->vlp[2] = NULL;
2211 	}
2212 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
2213 }
2214 
2215 static void
2216 cache_enter_unlock(struct celockstate *cel)
2217 {
2218 
2219 	cache_unlock_buckets_cel(cel);
2220 	cache_unlock_vnodes_cel(cel);
2221 }
2222 
2223 static void __noinline
2224 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp,
2225     struct componentname *cnp)
2226 {
2227 	struct celockstate cel;
2228 	struct namecache *ncp;
2229 	uint32_t hash;
2230 	int len;
2231 
2232 	if (atomic_load_ptr(&dvp->v_cache_dd) == NULL)
2233 		return;
2234 	len = cnp->cn_namelen;
2235 	cache_celockstate_init(&cel);
2236 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2237 	cache_enter_lock_dd(&cel, dvp, vp, hash);
2238 	ncp = dvp->v_cache_dd;
2239 	if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) {
2240 		KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent"));
2241 		cache_zap_locked(ncp);
2242 	} else {
2243 		ncp = NULL;
2244 	}
2245 	atomic_store_ptr(&dvp->v_cache_dd, NULL);
2246 	cache_enter_unlock(&cel);
2247 	if (ncp != NULL)
2248 		cache_free(ncp);
2249 }
2250 
2251 /*
2252  * Add an entry to the cache.
2253  */
2254 void
2255 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
2256     struct timespec *tsp, struct timespec *dtsp)
2257 {
2258 	struct celockstate cel;
2259 	struct namecache *ncp, *n2, *ndd;
2260 	struct namecache_ts *ncp_ts;
2261 	struct nchashhead *ncpp;
2262 	uint32_t hash;
2263 	int flag;
2264 	int len;
2265 
2266 	KASSERT(cnp->cn_namelen <= NAME_MAX,
2267 	    ("%s: passed len %ld exceeds NAME_MAX (%d)", __func__, cnp->cn_namelen,
2268 	    NAME_MAX));
2269 #ifdef notyet
2270 	/*
2271 	 * Not everything doing this is weeded out yet.
2272 	 */
2273 	VNPASS(dvp != vp, dvp);
2274 #endif
2275 	VNPASS(!VN_IS_DOOMED(dvp), dvp);
2276 	VNPASS(dvp->v_type != VNON, dvp);
2277 	if (vp != NULL) {
2278 		VNPASS(!VN_IS_DOOMED(vp), vp);
2279 		VNPASS(vp->v_type != VNON, vp);
2280 	}
2281 
2282 #ifdef DEBUG_CACHE
2283 	if (__predict_false(!doingcache))
2284 		return;
2285 #endif
2286 
2287 	flag = 0;
2288 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
2289 		if (cnp->cn_namelen == 1)
2290 			return;
2291 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
2292 			cache_enter_dotdot_prep(dvp, vp, cnp);
2293 			flag = NCF_ISDOTDOT;
2294 		}
2295 	}
2296 
2297 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
2298 	if (ncp == NULL)
2299 		return;
2300 
2301 	cache_celockstate_init(&cel);
2302 	ndd = NULL;
2303 	ncp_ts = NULL;
2304 
2305 	/*
2306 	 * Calculate the hash key and setup as much of the new
2307 	 * namecache entry as possible before acquiring the lock.
2308 	 */
2309 	ncp->nc_flag = flag | NCF_WIP;
2310 	ncp->nc_vp = vp;
2311 	if (vp == NULL)
2312 		cache_neg_init(ncp);
2313 	ncp->nc_dvp = dvp;
2314 	if (tsp != NULL) {
2315 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
2316 		ncp_ts->nc_time = *tsp;
2317 		ncp_ts->nc_ticks = ticks;
2318 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
2319 		if (dtsp != NULL) {
2320 			ncp_ts->nc_dotdottime = *dtsp;
2321 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
2322 		}
2323 	}
2324 	len = ncp->nc_nlen = cnp->cn_namelen;
2325 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2326 	memcpy(ncp->nc_name, cnp->cn_nameptr, len);
2327 	ncp->nc_name[len] = '\0';
2328 	cache_enter_lock(&cel, dvp, vp, hash);
2329 
2330 	/*
2331 	 * See if this vnode or negative entry is already in the cache
2332 	 * with this name.  This can happen with concurrent lookups of
2333 	 * the same path name.
2334 	 */
2335 	ncpp = NCHHASH(hash);
2336 	CK_SLIST_FOREACH(n2, ncpp, nc_hash) {
2337 		if (n2->nc_dvp == dvp &&
2338 		    n2->nc_nlen == cnp->cn_namelen &&
2339 		    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
2340 			MPASS(cache_ncp_canuse(n2));
2341 			if ((n2->nc_flag & NCF_NEGATIVE) != 0)
2342 				KASSERT(vp == NULL,
2343 				    ("%s: found entry pointing to a different vnode (%p != %p)",
2344 				    __func__, NULL, vp));
2345 			else
2346 				KASSERT(n2->nc_vp == vp,
2347 				    ("%s: found entry pointing to a different vnode (%p != %p)",
2348 				    __func__, n2->nc_vp, vp));
2349 			/*
2350 			 * Entries are supposed to be immutable unless in the
2351 			 * process of getting destroyed. Accommodating for
2352 			 * changing timestamps is possible but not worth it.
2353 			 * This should be harmless in terms of correctness, in
2354 			 * the worst case resulting in an earlier expiration.
2355 			 * Alternatively, the found entry can be replaced
2356 			 * altogether.
2357 			 */
2358 			MPASS((n2->nc_flag & (NCF_TS | NCF_DTS)) == (ncp->nc_flag & (NCF_TS | NCF_DTS)));
2359 #if 0
2360 			if (tsp != NULL) {
2361 				KASSERT((n2->nc_flag & NCF_TS) != 0,
2362 				    ("no NCF_TS"));
2363 				n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
2364 				n2_ts->nc_time = ncp_ts->nc_time;
2365 				n2_ts->nc_ticks = ncp_ts->nc_ticks;
2366 				if (dtsp != NULL) {
2367 					n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
2368 					n2_ts->nc_nc.nc_flag |= NCF_DTS;
2369 				}
2370 			}
2371 #endif
2372 			SDT_PROBE3(vfs, namecache, enter, duplicate, dvp, ncp->nc_name,
2373 			    vp);
2374 			goto out_unlock_free;
2375 		}
2376 	}
2377 
2378 	if (flag == NCF_ISDOTDOT) {
2379 		/*
2380 		 * See if we are trying to add .. entry, but some other lookup
2381 		 * has populated v_cache_dd pointer already.
2382 		 */
2383 		if (dvp->v_cache_dd != NULL)
2384 			goto out_unlock_free;
2385 		KASSERT(vp == NULL || vp->v_type == VDIR,
2386 		    ("wrong vnode type %p", vp));
2387 		atomic_thread_fence_rel();
2388 		atomic_store_ptr(&dvp->v_cache_dd, ncp);
2389 	}
2390 
2391 	if (vp != NULL) {
2392 		if (flag != NCF_ISDOTDOT) {
2393 			/*
2394 			 * For this case, the cache entry maps both the
2395 			 * directory name in it and the name ".." for the
2396 			 * directory's parent.
2397 			 */
2398 			if ((ndd = vp->v_cache_dd) != NULL) {
2399 				if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
2400 					cache_zap_locked(ndd);
2401 				else
2402 					ndd = NULL;
2403 			}
2404 			atomic_thread_fence_rel();
2405 			atomic_store_ptr(&vp->v_cache_dd, ncp);
2406 		} else if (vp->v_type != VDIR) {
2407 			if (vp->v_cache_dd != NULL) {
2408 				atomic_store_ptr(&vp->v_cache_dd, NULL);
2409 			}
2410 		}
2411 	}
2412 
2413 	if (flag != NCF_ISDOTDOT) {
2414 		if (LIST_EMPTY(&dvp->v_cache_src)) {
2415 			cache_hold_vnode(dvp);
2416 		}
2417 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
2418 	}
2419 
2420 	/*
2421 	 * If the entry is "negative", we place it into the
2422 	 * "negative" cache queue, otherwise, we place it into the
2423 	 * destination vnode's cache entries queue.
2424 	 */
2425 	if (vp != NULL) {
2426 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
2427 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
2428 		    vp);
2429 	} else {
2430 		if (cnp->cn_flags & ISWHITEOUT)
2431 			atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_WHITE);
2432 		cache_neg_insert(ncp);
2433 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
2434 		    ncp->nc_name);
2435 	}
2436 
2437 	/*
2438 	 * Insert the new namecache entry into the appropriate chain
2439 	 * within the cache entries table.
2440 	 */
2441 	CK_SLIST_INSERT_HEAD(ncpp, ncp, nc_hash);
2442 
2443 	atomic_thread_fence_rel();
2444 	/*
2445 	 * Mark the entry as fully constructed.
2446 	 * It is immutable past this point until its removal.
2447 	 */
2448 	atomic_store_char(&ncp->nc_flag, ncp->nc_flag & ~NCF_WIP);
2449 
2450 	cache_enter_unlock(&cel);
2451 	if (ndd != NULL)
2452 		cache_free(ndd);
2453 	return;
2454 out_unlock_free:
2455 	cache_enter_unlock(&cel);
2456 	cache_free(ncp);
2457 	return;
2458 }
2459 
2460 static u_int
2461 cache_roundup_2(u_int val)
2462 {
2463 	u_int res;
2464 
2465 	for (res = 1; res <= val; res <<= 1)
2466 		continue;
2467 
2468 	return (res);
2469 }
2470 
2471 static struct nchashhead *
2472 nchinittbl(u_long elements, u_long *hashmask)
2473 {
2474 	struct nchashhead *hashtbl;
2475 	u_long hashsize, i;
2476 
2477 	hashsize = cache_roundup_2(elements) / 2;
2478 
2479 	hashtbl = malloc((u_long)hashsize * sizeof(*hashtbl), M_VFSCACHE, M_WAITOK);
2480 	for (i = 0; i < hashsize; i++)
2481 		CK_SLIST_INIT(&hashtbl[i]);
2482 	*hashmask = hashsize - 1;
2483 	return (hashtbl);
2484 }
2485 
2486 static void
2487 ncfreetbl(struct nchashhead *hashtbl)
2488 {
2489 
2490 	free(hashtbl, M_VFSCACHE);
2491 }
2492 
2493 /*
2494  * Name cache initialization, from vfs_init() when we are booting
2495  */
2496 static void
2497 nchinit(void *dummy __unused)
2498 {
2499 	u_int i;
2500 
2501 	cache_zone_small = uma_zcreate("S VFS Cache", CACHE_ZONE_SMALL_SIZE,
2502 	    NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2503 	cache_zone_small_ts = uma_zcreate("STS VFS Cache", CACHE_ZONE_SMALL_TS_SIZE,
2504 	    NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2505 	cache_zone_large = uma_zcreate("L VFS Cache", CACHE_ZONE_LARGE_SIZE,
2506 	    NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2507 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache", CACHE_ZONE_LARGE_TS_SIZE,
2508 	    NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2509 
2510 	VFS_SMR_ZONE_SET(cache_zone_small);
2511 	VFS_SMR_ZONE_SET(cache_zone_small_ts);
2512 	VFS_SMR_ZONE_SET(cache_zone_large);
2513 	VFS_SMR_ZONE_SET(cache_zone_large_ts);
2514 
2515 	ncsize = desiredvnodes * ncsizefactor;
2516 	cache_recalc_neg_min(ncnegminpct);
2517 	nchashtbl = nchinittbl(desiredvnodes * 2, &nchash);
2518 	ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
2519 	if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
2520 		ncbuckethash = 7;
2521 	if (ncbuckethash > nchash)
2522 		ncbuckethash = nchash;
2523 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
2524 	    M_WAITOK | M_ZERO);
2525 	for (i = 0; i < numbucketlocks; i++)
2526 		mtx_init(&bucketlocks[i], "ncbuc", NULL, MTX_DUPOK | MTX_RECURSE);
2527 	ncvnodehash = ncbuckethash;
2528 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
2529 	    M_WAITOK | M_ZERO);
2530 	for (i = 0; i < numvnodelocks; i++)
2531 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
2532 
2533 	for (i = 0; i < numneglists; i++) {
2534 		mtx_init(&neglists[i].nl_evict_lock, "ncnege", NULL, MTX_DEF);
2535 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
2536 		TAILQ_INIT(&neglists[i].nl_list);
2537 		TAILQ_INIT(&neglists[i].nl_hotlist);
2538 	}
2539 }
2540 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
2541 
2542 void
2543 cache_vnode_init(struct vnode *vp)
2544 {
2545 
2546 	LIST_INIT(&vp->v_cache_src);
2547 	TAILQ_INIT(&vp->v_cache_dst);
2548 	vp->v_cache_dd = NULL;
2549 	cache_prehash(vp);
2550 }
2551 
2552 void
2553 cache_changesize(u_long newmaxvnodes)
2554 {
2555 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
2556 	u_long new_nchash, old_nchash;
2557 	struct namecache *ncp;
2558 	uint32_t hash;
2559 	u_long newncsize;
2560 	int i;
2561 
2562 	newncsize = newmaxvnodes * ncsizefactor;
2563 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
2564 	if (newmaxvnodes < numbucketlocks)
2565 		newmaxvnodes = numbucketlocks;
2566 
2567 	new_nchashtbl = nchinittbl(newmaxvnodes, &new_nchash);
2568 	/* If same hash table size, nothing to do */
2569 	if (nchash == new_nchash) {
2570 		ncfreetbl(new_nchashtbl);
2571 		return;
2572 	}
2573 	/*
2574 	 * Move everything from the old hash table to the new table.
2575 	 * None of the namecache entries in the table can be removed
2576 	 * because to do so, they have to be removed from the hash table.
2577 	 */
2578 	cache_lock_all_vnodes();
2579 	cache_lock_all_buckets();
2580 	old_nchashtbl = nchashtbl;
2581 	old_nchash = nchash;
2582 	nchashtbl = new_nchashtbl;
2583 	nchash = new_nchash;
2584 	for (i = 0; i <= old_nchash; i++) {
2585 		while ((ncp = CK_SLIST_FIRST(&old_nchashtbl[i])) != NULL) {
2586 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
2587 			    ncp->nc_dvp);
2588 			CK_SLIST_REMOVE(&old_nchashtbl[i], ncp, namecache, nc_hash);
2589 			CK_SLIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
2590 		}
2591 	}
2592 	ncsize = newncsize;
2593 	cache_recalc_neg_min(ncnegminpct);
2594 	cache_unlock_all_buckets();
2595 	cache_unlock_all_vnodes();
2596 	ncfreetbl(old_nchashtbl);
2597 }
2598 
2599 /*
2600  * Remove all entries from and to a particular vnode.
2601  */
2602 static void
2603 cache_purge_impl(struct vnode *vp)
2604 {
2605 	struct cache_freebatch batch;
2606 	struct namecache *ncp;
2607 	struct mtx *vlp, *vlp2;
2608 
2609 	TAILQ_INIT(&batch);
2610 	vlp = VP2VNODELOCK(vp);
2611 	vlp2 = NULL;
2612 	mtx_lock(vlp);
2613 retry:
2614 	while (!LIST_EMPTY(&vp->v_cache_src)) {
2615 		ncp = LIST_FIRST(&vp->v_cache_src);
2616 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2617 			goto retry;
2618 		TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2619 	}
2620 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2621 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
2622 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2623 			goto retry;
2624 		TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2625 	}
2626 	ncp = vp->v_cache_dd;
2627 	if (ncp != NULL) {
2628 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2629 		   ("lost dotdot link"));
2630 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2631 			goto retry;
2632 		TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2633 	}
2634 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2635 	mtx_unlock(vlp);
2636 	if (vlp2 != NULL)
2637 		mtx_unlock(vlp2);
2638 	cache_free_batch(&batch);
2639 }
2640 
2641 /*
2642  * Opportunistic check to see if there is anything to do.
2643  */
2644 static bool
2645 cache_has_entries(struct vnode *vp)
2646 {
2647 
2648 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2649 	    atomic_load_ptr(&vp->v_cache_dd) == NULL)
2650 		return (false);
2651 	return (true);
2652 }
2653 
2654 void
2655 cache_purge(struct vnode *vp)
2656 {
2657 
2658 	SDT_PROBE1(vfs, namecache, purge, done, vp);
2659 	if (!cache_has_entries(vp))
2660 		return;
2661 	cache_purge_impl(vp);
2662 }
2663 
2664 /*
2665  * Only to be used by vgone.
2666  */
2667 void
2668 cache_purge_vgone(struct vnode *vp)
2669 {
2670 	struct mtx *vlp;
2671 
2672 	VNPASS(VN_IS_DOOMED(vp), vp);
2673 	if (cache_has_entries(vp)) {
2674 		cache_purge_impl(vp);
2675 		return;
2676 	}
2677 
2678 	/*
2679 	 * Serialize against a potential thread doing cache_purge.
2680 	 */
2681 	vlp = VP2VNODELOCK(vp);
2682 	mtx_wait_unlocked(vlp);
2683 	if (cache_has_entries(vp)) {
2684 		cache_purge_impl(vp);
2685 		return;
2686 	}
2687 	return;
2688 }
2689 
2690 /*
2691  * Remove all negative entries for a particular directory vnode.
2692  */
2693 void
2694 cache_purge_negative(struct vnode *vp)
2695 {
2696 	struct cache_freebatch batch;
2697 	struct namecache *ncp, *nnp;
2698 	struct mtx *vlp;
2699 
2700 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2701 	if (LIST_EMPTY(&vp->v_cache_src))
2702 		return;
2703 	TAILQ_INIT(&batch);
2704 	vlp = VP2VNODELOCK(vp);
2705 	mtx_lock(vlp);
2706 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2707 		if (!(ncp->nc_flag & NCF_NEGATIVE))
2708 			continue;
2709 		cache_zap_negative_locked_vnode_kl(ncp, vp);
2710 		TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2711 	}
2712 	mtx_unlock(vlp);
2713 	cache_free_batch(&batch);
2714 }
2715 
2716 /*
2717  * Entry points for modifying VOP operations.
2718  */
2719 void
2720 cache_vop_rename(struct vnode *fdvp, struct vnode *fvp, struct vnode *tdvp,
2721     struct vnode *tvp, struct componentname *fcnp, struct componentname *tcnp)
2722 {
2723 
2724 	ASSERT_VOP_IN_SEQC(fdvp);
2725 	ASSERT_VOP_IN_SEQC(fvp);
2726 	ASSERT_VOP_IN_SEQC(tdvp);
2727 	if (tvp != NULL)
2728 		ASSERT_VOP_IN_SEQC(tvp);
2729 
2730 	cache_purge(fvp);
2731 	if (tvp != NULL) {
2732 		cache_purge(tvp);
2733 		KASSERT(!cache_remove_cnp(tdvp, tcnp),
2734 		    ("%s: lingering negative entry", __func__));
2735 	} else {
2736 		cache_remove_cnp(tdvp, tcnp);
2737 	}
2738 
2739 	/*
2740 	 * TODO
2741 	 *
2742 	 * Historically renaming was always purging all revelang entries,
2743 	 * but that's quite wasteful. In particular turns out that in many cases
2744 	 * the target file is immediately accessed after rename, inducing a cache
2745 	 * miss.
2746 	 *
2747 	 * Recode this to reduce relocking and reuse the existing entry (if any)
2748 	 * instead of just removing it above and allocating a new one here.
2749 	 */
2750 	if (cache_rename_add) {
2751 		cache_enter(tdvp, fvp, tcnp);
2752 	}
2753 }
2754 
2755 void
2756 cache_vop_rmdir(struct vnode *dvp, struct vnode *vp)
2757 {
2758 
2759 	ASSERT_VOP_IN_SEQC(dvp);
2760 	ASSERT_VOP_IN_SEQC(vp);
2761 	cache_purge(vp);
2762 }
2763 
2764 #ifdef INVARIANTS
2765 /*
2766  * Validate that if an entry exists it matches.
2767  */
2768 void
2769 cache_validate(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
2770 {
2771 	struct namecache *ncp;
2772 	struct mtx *blp;
2773 	uint32_t hash;
2774 
2775 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
2776 	if (CK_SLIST_EMPTY(NCHHASH(hash)))
2777 		return;
2778 	blp = HASH2BUCKETLOCK(hash);
2779 	mtx_lock(blp);
2780 	CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
2781 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
2782 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) {
2783 			if (ncp->nc_vp != vp)
2784 				panic("%s: mismatch (%p != %p); ncp %p [%s] dvp %p\n",
2785 				    __func__, vp, ncp->nc_vp, ncp, ncp->nc_name, ncp->nc_dvp);
2786 		}
2787 	}
2788 	mtx_unlock(blp);
2789 }
2790 #endif
2791 
2792 /*
2793  * Flush all entries referencing a particular filesystem.
2794  */
2795 void
2796 cache_purgevfs(struct mount *mp)
2797 {
2798 	struct vnode *vp, *mvp;
2799 
2800 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2801 	/*
2802 	 * Somewhat wasteful iteration over all vnodes. Would be better to
2803 	 * support filtering and avoid the interlock to begin with.
2804 	 */
2805 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2806 		if (!cache_has_entries(vp)) {
2807 			VI_UNLOCK(vp);
2808 			continue;
2809 		}
2810 		vholdl(vp);
2811 		VI_UNLOCK(vp);
2812 		cache_purge(vp);
2813 		vdrop(vp);
2814 	}
2815 }
2816 
2817 /*
2818  * Perform canonical checks and cache lookup and pass on to filesystem
2819  * through the vop_cachedlookup only if needed.
2820  */
2821 
2822 int
2823 vfs_cache_lookup(struct vop_lookup_args *ap)
2824 {
2825 	struct vnode *dvp;
2826 	int error;
2827 	struct vnode **vpp = ap->a_vpp;
2828 	struct componentname *cnp = ap->a_cnp;
2829 	int flags = cnp->cn_flags;
2830 
2831 	*vpp = NULL;
2832 	dvp = ap->a_dvp;
2833 
2834 	if (dvp->v_type != VDIR)
2835 		return (ENOTDIR);
2836 
2837 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2838 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2839 		return (EROFS);
2840 
2841 	error = vn_dir_check_exec(dvp, cnp);
2842 	if (error != 0)
2843 		return (error);
2844 
2845 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2846 	if (error == 0)
2847 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2848 	if (error == -1)
2849 		return (0);
2850 	return (error);
2851 }
2852 
2853 /* Implementation of the getcwd syscall. */
2854 int
2855 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2856 {
2857 	char *buf, *retbuf;
2858 	size_t buflen;
2859 	int error;
2860 
2861 	buflen = uap->buflen;
2862 	if (__predict_false(buflen < 2))
2863 		return (EINVAL);
2864 	if (buflen > MAXPATHLEN)
2865 		buflen = MAXPATHLEN;
2866 
2867 	buf = uma_zalloc(namei_zone, M_WAITOK);
2868 	error = vn_getcwd(buf, &retbuf, &buflen);
2869 	if (error == 0)
2870 		error = copyout(retbuf, uap->buf, buflen);
2871 	uma_zfree(namei_zone, buf);
2872 	return (error);
2873 }
2874 
2875 int
2876 vn_getcwd(char *buf, char **retbuf, size_t *buflen)
2877 {
2878 	struct pwd *pwd;
2879 	int error;
2880 
2881 	vfs_smr_enter();
2882 	pwd = pwd_get_smr();
2883 	error = vn_fullpath_any_smr(pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf,
2884 	    buflen, 0);
2885 	VFS_SMR_ASSERT_NOT_ENTERED();
2886 	if (error < 0) {
2887 		pwd = pwd_hold(curthread);
2888 		error = vn_fullpath_any(pwd->pwd_cdir, pwd->pwd_rdir, buf,
2889 		    retbuf, buflen);
2890 		pwd_drop(pwd);
2891 	}
2892 
2893 #ifdef KTRACE
2894 	if (KTRPOINT(curthread, KTR_NAMEI) && error == 0)
2895 		ktrnamei(*retbuf);
2896 #endif
2897 	return (error);
2898 }
2899 
2900 static int
2901 kern___realpathat(struct thread *td, int fd, const char *path, char *buf,
2902     size_t size, int flags, enum uio_seg pathseg)
2903 {
2904 	struct nameidata nd;
2905 	char *retbuf, *freebuf;
2906 	int error;
2907 
2908 	if (flags != 0)
2909 		return (EINVAL);
2910 	NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | SAVENAME | WANTPARENT | AUDITVNODE1,
2911 	    pathseg, path, fd, &cap_fstat_rights, td);
2912 	if ((error = namei(&nd)) != 0)
2913 		return (error);
2914 	error = vn_fullpath_hardlink(&nd, &retbuf, &freebuf, &size);
2915 	if (error == 0) {
2916 		error = copyout(retbuf, buf, size);
2917 		free(freebuf, M_TEMP);
2918 	}
2919 	NDFREE(&nd, 0);
2920 	return (error);
2921 }
2922 
2923 int
2924 sys___realpathat(struct thread *td, struct __realpathat_args *uap)
2925 {
2926 
2927 	return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size,
2928 	    uap->flags, UIO_USERSPACE));
2929 }
2930 
2931 /*
2932  * Retrieve the full filesystem path that correspond to a vnode from the name
2933  * cache (if available)
2934  */
2935 int
2936 vn_fullpath(struct vnode *vp, char **retbuf, char **freebuf)
2937 {
2938 	struct pwd *pwd;
2939 	char *buf;
2940 	size_t buflen;
2941 	int error;
2942 
2943 	if (__predict_false(vp == NULL))
2944 		return (EINVAL);
2945 
2946 	buflen = MAXPATHLEN;
2947 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2948 	vfs_smr_enter();
2949 	pwd = pwd_get_smr();
2950 	error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, &buflen, 0);
2951 	VFS_SMR_ASSERT_NOT_ENTERED();
2952 	if (error < 0) {
2953 		pwd = pwd_hold(curthread);
2954 		error = vn_fullpath_any(vp, pwd->pwd_rdir, buf, retbuf, &buflen);
2955 		pwd_drop(pwd);
2956 	}
2957 	if (error == 0)
2958 		*freebuf = buf;
2959 	else
2960 		free(buf, M_TEMP);
2961 	return (error);
2962 }
2963 
2964 /*
2965  * This function is similar to vn_fullpath, but it attempts to lookup the
2966  * pathname relative to the global root mount point.  This is required for the
2967  * auditing sub-system, as audited pathnames must be absolute, relative to the
2968  * global root mount point.
2969  */
2970 int
2971 vn_fullpath_global(struct vnode *vp, char **retbuf, char **freebuf)
2972 {
2973 	char *buf;
2974 	size_t buflen;
2975 	int error;
2976 
2977 	if (__predict_false(vp == NULL))
2978 		return (EINVAL);
2979 	buflen = MAXPATHLEN;
2980 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2981 	vfs_smr_enter();
2982 	error = vn_fullpath_any_smr(vp, rootvnode, buf, retbuf, &buflen, 0);
2983 	VFS_SMR_ASSERT_NOT_ENTERED();
2984 	if (error < 0) {
2985 		error = vn_fullpath_any(vp, rootvnode, buf, retbuf, &buflen);
2986 	}
2987 	if (error == 0)
2988 		*freebuf = buf;
2989 	else
2990 		free(buf, M_TEMP);
2991 	return (error);
2992 }
2993 
2994 static struct namecache *
2995 vn_dd_from_dst(struct vnode *vp)
2996 {
2997 	struct namecache *ncp;
2998 
2999 	cache_assert_vnode_locked(vp);
3000 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) {
3001 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
3002 			return (ncp);
3003 	}
3004 	return (NULL);
3005 }
3006 
3007 int
3008 vn_vptocnp(struct vnode **vp, char *buf, size_t *buflen)
3009 {
3010 	struct vnode *dvp;
3011 	struct namecache *ncp;
3012 	struct mtx *vlp;
3013 	int error;
3014 
3015 	vlp = VP2VNODELOCK(*vp);
3016 	mtx_lock(vlp);
3017 	ncp = (*vp)->v_cache_dd;
3018 	if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT) == 0) {
3019 		KASSERT(ncp == vn_dd_from_dst(*vp),
3020 		    ("%s: mismatch for dd entry (%p != %p)", __func__,
3021 		    ncp, vn_dd_from_dst(*vp)));
3022 	} else {
3023 		ncp = vn_dd_from_dst(*vp);
3024 	}
3025 	if (ncp != NULL) {
3026 		if (*buflen < ncp->nc_nlen) {
3027 			mtx_unlock(vlp);
3028 			vrele(*vp);
3029 			counter_u64_add(numfullpathfail4, 1);
3030 			error = ENOMEM;
3031 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
3032 			    vp, NULL);
3033 			return (error);
3034 		}
3035 		*buflen -= ncp->nc_nlen;
3036 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
3037 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
3038 		    ncp->nc_name, vp);
3039 		dvp = *vp;
3040 		*vp = ncp->nc_dvp;
3041 		vref(*vp);
3042 		mtx_unlock(vlp);
3043 		vrele(dvp);
3044 		return (0);
3045 	}
3046 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
3047 
3048 	mtx_unlock(vlp);
3049 	vn_lock(*vp, LK_SHARED | LK_RETRY);
3050 	error = VOP_VPTOCNP(*vp, &dvp, buf, buflen);
3051 	vput(*vp);
3052 	if (error) {
3053 		counter_u64_add(numfullpathfail2, 1);
3054 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
3055 		return (error);
3056 	}
3057 
3058 	*vp = dvp;
3059 	if (VN_IS_DOOMED(dvp)) {
3060 		/* forced unmount */
3061 		vrele(dvp);
3062 		error = ENOENT;
3063 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
3064 		return (error);
3065 	}
3066 	/*
3067 	 * *vp has its use count incremented still.
3068 	 */
3069 
3070 	return (0);
3071 }
3072 
3073 /*
3074  * Resolve a directory to a pathname.
3075  *
3076  * The name of the directory can always be found in the namecache or fetched
3077  * from the filesystem. There is also guaranteed to be only one parent, meaning
3078  * we can just follow vnodes up until we find the root.
3079  *
3080  * The vnode must be referenced.
3081  */
3082 static int
3083 vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3084     size_t *len, size_t addend)
3085 {
3086 #ifdef KDTRACE_HOOKS
3087 	struct vnode *startvp = vp;
3088 #endif
3089 	struct vnode *vp1;
3090 	size_t buflen;
3091 	int error;
3092 	bool slash_prefixed;
3093 
3094 	VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp);
3095 	VNPASS(vp->v_usecount > 0, vp);
3096 
3097 	buflen = *len;
3098 
3099 	slash_prefixed = true;
3100 	if (addend == 0) {
3101 		MPASS(*len >= 2);
3102 		buflen--;
3103 		buf[buflen] = '\0';
3104 		slash_prefixed = false;
3105 	}
3106 
3107 	error = 0;
3108 
3109 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
3110 	counter_u64_add(numfullpathcalls, 1);
3111 	while (vp != rdir && vp != rootvnode) {
3112 		/*
3113 		 * The vp vnode must be already fully constructed,
3114 		 * since it is either found in namecache or obtained
3115 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
3116 		 * without obtaining the vnode lock.
3117 		 */
3118 		if ((vp->v_vflag & VV_ROOT) != 0) {
3119 			vn_lock(vp, LK_RETRY | LK_SHARED);
3120 
3121 			/*
3122 			 * With the vnode locked, check for races with
3123 			 * unmount, forced or not.  Note that we
3124 			 * already verified that vp is not equal to
3125 			 * the root vnode, which means that
3126 			 * mnt_vnodecovered can be NULL only for the
3127 			 * case of unmount.
3128 			 */
3129 			if (VN_IS_DOOMED(vp) ||
3130 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
3131 			    vp1->v_mountedhere != vp->v_mount) {
3132 				vput(vp);
3133 				error = ENOENT;
3134 				SDT_PROBE3(vfs, namecache, fullpath, return,
3135 				    error, vp, NULL);
3136 				break;
3137 			}
3138 
3139 			vref(vp1);
3140 			vput(vp);
3141 			vp = vp1;
3142 			continue;
3143 		}
3144 		if (vp->v_type != VDIR) {
3145 			vrele(vp);
3146 			counter_u64_add(numfullpathfail1, 1);
3147 			error = ENOTDIR;
3148 			SDT_PROBE3(vfs, namecache, fullpath, return,
3149 			    error, vp, NULL);
3150 			break;
3151 		}
3152 		error = vn_vptocnp(&vp, buf, &buflen);
3153 		if (error)
3154 			break;
3155 		if (buflen == 0) {
3156 			vrele(vp);
3157 			error = ENOMEM;
3158 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
3159 			    startvp, NULL);
3160 			break;
3161 		}
3162 		buf[--buflen] = '/';
3163 		slash_prefixed = true;
3164 	}
3165 	if (error)
3166 		return (error);
3167 	if (!slash_prefixed) {
3168 		if (buflen == 0) {
3169 			vrele(vp);
3170 			counter_u64_add(numfullpathfail4, 1);
3171 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
3172 			    startvp, NULL);
3173 			return (ENOMEM);
3174 		}
3175 		buf[--buflen] = '/';
3176 	}
3177 	counter_u64_add(numfullpathfound, 1);
3178 	vrele(vp);
3179 
3180 	*retbuf = buf + buflen;
3181 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf);
3182 	*len -= buflen;
3183 	*len += addend;
3184 	return (0);
3185 }
3186 
3187 /*
3188  * Resolve an arbitrary vnode to a pathname.
3189  *
3190  * Note 2 caveats:
3191  * - hardlinks are not tracked, thus if the vnode is not a directory this can
3192  *   resolve to a different path than the one used to find it
3193  * - namecache is not mandatory, meaning names are not guaranteed to be added
3194  *   (in which case resolving fails)
3195  */
3196 static void __inline
3197 cache_rev_failed_impl(int *reason, int line)
3198 {
3199 
3200 	*reason = line;
3201 }
3202 #define cache_rev_failed(var)	cache_rev_failed_impl((var), __LINE__)
3203 
3204 static int
3205 vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
3206     char **retbuf, size_t *buflen, size_t addend)
3207 {
3208 #ifdef KDTRACE_HOOKS
3209 	struct vnode *startvp = vp;
3210 #endif
3211 	struct vnode *tvp;
3212 	struct mount *mp;
3213 	struct namecache *ncp;
3214 	size_t orig_buflen;
3215 	int reason;
3216 	int error;
3217 #ifdef KDTRACE_HOOKS
3218 	int i;
3219 #endif
3220 	seqc_t vp_seqc, tvp_seqc;
3221 	u_char nc_flag;
3222 
3223 	VFS_SMR_ASSERT_ENTERED();
3224 
3225 	if (!cache_fast_revlookup) {
3226 		vfs_smr_exit();
3227 		return (-1);
3228 	}
3229 
3230 	orig_buflen = *buflen;
3231 
3232 	if (addend == 0) {
3233 		MPASS(*buflen >= 2);
3234 		*buflen -= 1;
3235 		buf[*buflen] = '\0';
3236 	}
3237 
3238 	if (vp == rdir || vp == rootvnode) {
3239 		if (addend == 0) {
3240 			*buflen -= 1;
3241 			buf[*buflen] = '/';
3242 		}
3243 		goto out_ok;
3244 	}
3245 
3246 #ifdef KDTRACE_HOOKS
3247 	i = 0;
3248 #endif
3249 	error = -1;
3250 	ncp = NULL; /* for sdt probe down below */
3251 	vp_seqc = vn_seqc_read_any(vp);
3252 	if (seqc_in_modify(vp_seqc)) {
3253 		cache_rev_failed(&reason);
3254 		goto out_abort;
3255 	}
3256 
3257 	for (;;) {
3258 #ifdef KDTRACE_HOOKS
3259 		i++;
3260 #endif
3261 		if ((vp->v_vflag & VV_ROOT) != 0) {
3262 			mp = atomic_load_ptr(&vp->v_mount);
3263 			if (mp == NULL) {
3264 				cache_rev_failed(&reason);
3265 				goto out_abort;
3266 			}
3267 			tvp = atomic_load_ptr(&mp->mnt_vnodecovered);
3268 			tvp_seqc = vn_seqc_read_any(tvp);
3269 			if (seqc_in_modify(tvp_seqc)) {
3270 				cache_rev_failed(&reason);
3271 				goto out_abort;
3272 			}
3273 			if (!vn_seqc_consistent(vp, vp_seqc)) {
3274 				cache_rev_failed(&reason);
3275 				goto out_abort;
3276 			}
3277 			vp = tvp;
3278 			vp_seqc = tvp_seqc;
3279 			continue;
3280 		}
3281 		ncp = atomic_load_consume_ptr(&vp->v_cache_dd);
3282 		if (ncp == NULL) {
3283 			cache_rev_failed(&reason);
3284 			goto out_abort;
3285 		}
3286 		nc_flag = atomic_load_char(&ncp->nc_flag);
3287 		if ((nc_flag & NCF_ISDOTDOT) != 0) {
3288 			cache_rev_failed(&reason);
3289 			goto out_abort;
3290 		}
3291 		if (ncp->nc_nlen >= *buflen) {
3292 			cache_rev_failed(&reason);
3293 			error = ENOMEM;
3294 			goto out_abort;
3295 		}
3296 		*buflen -= ncp->nc_nlen;
3297 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
3298 		*buflen -= 1;
3299 		buf[*buflen] = '/';
3300 		tvp = ncp->nc_dvp;
3301 		tvp_seqc = vn_seqc_read_any(tvp);
3302 		if (seqc_in_modify(tvp_seqc)) {
3303 			cache_rev_failed(&reason);
3304 			goto out_abort;
3305 		}
3306 		if (!vn_seqc_consistent(vp, vp_seqc)) {
3307 			cache_rev_failed(&reason);
3308 			goto out_abort;
3309 		}
3310 		/*
3311 		 * Acquire fence provided by vn_seqc_read_any above.
3312 		 */
3313 		if (__predict_false(atomic_load_ptr(&vp->v_cache_dd) != ncp)) {
3314 			cache_rev_failed(&reason);
3315 			goto out_abort;
3316 		}
3317 		if (!cache_ncp_canuse(ncp)) {
3318 			cache_rev_failed(&reason);
3319 			goto out_abort;
3320 		}
3321 		vp = tvp;
3322 		vp_seqc = tvp_seqc;
3323 		if (vp == rdir || vp == rootvnode)
3324 			break;
3325 	}
3326 out_ok:
3327 	vfs_smr_exit();
3328 	*retbuf = buf + *buflen;
3329 	*buflen = orig_buflen - *buflen + addend;
3330 	SDT_PROBE2(vfs, namecache, fullpath_smr, hit, startvp, *retbuf);
3331 	return (0);
3332 
3333 out_abort:
3334 	*buflen = orig_buflen;
3335 	SDT_PROBE4(vfs, namecache, fullpath_smr, miss, startvp, ncp, reason, i);
3336 	vfs_smr_exit();
3337 	return (error);
3338 }
3339 
3340 static int
3341 vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3342     size_t *buflen)
3343 {
3344 	size_t orig_buflen, addend;
3345 	int error;
3346 
3347 	if (*buflen < 2)
3348 		return (EINVAL);
3349 
3350 	orig_buflen = *buflen;
3351 
3352 	vref(vp);
3353 	addend = 0;
3354 	if (vp->v_type != VDIR) {
3355 		*buflen -= 1;
3356 		buf[*buflen] = '\0';
3357 		error = vn_vptocnp(&vp, buf, buflen);
3358 		if (error)
3359 			return (error);
3360 		if (*buflen == 0) {
3361 			vrele(vp);
3362 			return (ENOMEM);
3363 		}
3364 		*buflen -= 1;
3365 		buf[*buflen] = '/';
3366 		addend = orig_buflen - *buflen;
3367 	}
3368 
3369 	return (vn_fullpath_dir(vp, rdir, buf, retbuf, buflen, addend));
3370 }
3371 
3372 /*
3373  * Resolve an arbitrary vnode to a pathname (taking care of hardlinks).
3374  *
3375  * Since the namecache does not track hardlinks, the caller is expected to first
3376  * look up the target vnode with SAVENAME | WANTPARENT flags passed to namei.
3377  *
3378  * Then we have 2 cases:
3379  * - if the found vnode is a directory, the path can be constructed just by
3380  *   following names up the chain
3381  * - otherwise we populate the buffer with the saved name and start resolving
3382  *   from the parent
3383  */
3384 static int
3385 vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf, char **freebuf,
3386     size_t *buflen)
3387 {
3388 	char *buf, *tmpbuf;
3389 	struct pwd *pwd;
3390 	struct componentname *cnp;
3391 	struct vnode *vp;
3392 	size_t addend;
3393 	int error;
3394 	enum vtype type;
3395 
3396 	if (*buflen < 2)
3397 		return (EINVAL);
3398 	if (*buflen > MAXPATHLEN)
3399 		*buflen = MAXPATHLEN;
3400 
3401 	buf = malloc(*buflen, M_TEMP, M_WAITOK);
3402 
3403 	addend = 0;
3404 	vp = ndp->ni_vp;
3405 	/*
3406 	 * Check for VBAD to work around the vp_crossmp bug in lookup().
3407 	 *
3408 	 * For example consider tmpfs on /tmp and realpath /tmp. ni_vp will be
3409 	 * set to mount point's root vnode while ni_dvp will be vp_crossmp.
3410 	 * If the type is VDIR (like in this very case) we can skip looking
3411 	 * at ni_dvp in the first place. However, since vnodes get passed here
3412 	 * unlocked the target may transition to doomed state (type == VBAD)
3413 	 * before we get to evaluate the condition. If this happens, we will
3414 	 * populate part of the buffer and descend to vn_fullpath_dir with
3415 	 * vp == vp_crossmp. Prevent the problem by checking for VBAD.
3416 	 *
3417 	 * This should be atomic_load(&vp->v_type) but it is illegal to take
3418 	 * an address of a bit field, even if said field is sized to char.
3419 	 * Work around the problem by reading the value into a full-sized enum
3420 	 * and then re-reading it with atomic_load which will still prevent
3421 	 * the compiler from re-reading down the road.
3422 	 */
3423 	type = vp->v_type;
3424 	type = atomic_load_int(&type);
3425 	if (type == VBAD) {
3426 		error = ENOENT;
3427 		goto out_bad;
3428 	}
3429 	if (type != VDIR) {
3430 		cnp = &ndp->ni_cnd;
3431 		addend = cnp->cn_namelen + 2;
3432 		if (*buflen < addend) {
3433 			error = ENOMEM;
3434 			goto out_bad;
3435 		}
3436 		*buflen -= addend;
3437 		tmpbuf = buf + *buflen;
3438 		tmpbuf[0] = '/';
3439 		memcpy(&tmpbuf[1], cnp->cn_nameptr, cnp->cn_namelen);
3440 		tmpbuf[addend - 1] = '\0';
3441 		vp = ndp->ni_dvp;
3442 	}
3443 
3444 	vfs_smr_enter();
3445 	pwd = pwd_get_smr();
3446 	error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3447 	    addend);
3448 	VFS_SMR_ASSERT_NOT_ENTERED();
3449 	if (error < 0) {
3450 		pwd = pwd_hold(curthread);
3451 		vref(vp);
3452 		error = vn_fullpath_dir(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3453 		    addend);
3454 		pwd_drop(pwd);
3455 		if (error != 0)
3456 			goto out_bad;
3457 	}
3458 
3459 	*freebuf = buf;
3460 
3461 	return (0);
3462 out_bad:
3463 	free(buf, M_TEMP);
3464 	return (error);
3465 }
3466 
3467 struct vnode *
3468 vn_dir_dd_ino(struct vnode *vp)
3469 {
3470 	struct namecache *ncp;
3471 	struct vnode *ddvp;
3472 	struct mtx *vlp;
3473 	enum vgetstate vs;
3474 
3475 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
3476 	vlp = VP2VNODELOCK(vp);
3477 	mtx_lock(vlp);
3478 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
3479 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
3480 			continue;
3481 		ddvp = ncp->nc_dvp;
3482 		vs = vget_prep(ddvp);
3483 		mtx_unlock(vlp);
3484 		if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
3485 			return (NULL);
3486 		return (ddvp);
3487 	}
3488 	mtx_unlock(vlp);
3489 	return (NULL);
3490 }
3491 
3492 int
3493 vn_commname(struct vnode *vp, char *buf, u_int buflen)
3494 {
3495 	struct namecache *ncp;
3496 	struct mtx *vlp;
3497 	int l;
3498 
3499 	vlp = VP2VNODELOCK(vp);
3500 	mtx_lock(vlp);
3501 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
3502 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
3503 			break;
3504 	if (ncp == NULL) {
3505 		mtx_unlock(vlp);
3506 		return (ENOENT);
3507 	}
3508 	l = min(ncp->nc_nlen, buflen - 1);
3509 	memcpy(buf, ncp->nc_name, l);
3510 	mtx_unlock(vlp);
3511 	buf[l] = '\0';
3512 	return (0);
3513 }
3514 
3515 /*
3516  * This function updates path string to vnode's full global path
3517  * and checks the size of the new path string against the pathlen argument.
3518  *
3519  * Requires a locked, referenced vnode.
3520  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
3521  *
3522  * If vp is a directory, the call to vn_fullpath_global() always succeeds
3523  * because it falls back to the ".." lookup if the namecache lookup fails.
3524  */
3525 int
3526 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
3527     u_int pathlen)
3528 {
3529 	struct nameidata nd;
3530 	struct vnode *vp1;
3531 	char *rpath, *fbuf;
3532 	int error;
3533 
3534 	ASSERT_VOP_ELOCKED(vp, __func__);
3535 
3536 	/* Construct global filesystem path from vp. */
3537 	VOP_UNLOCK(vp);
3538 	error = vn_fullpath_global(vp, &rpath, &fbuf);
3539 
3540 	if (error != 0) {
3541 		vrele(vp);
3542 		return (error);
3543 	}
3544 
3545 	if (strlen(rpath) >= pathlen) {
3546 		vrele(vp);
3547 		error = ENAMETOOLONG;
3548 		goto out;
3549 	}
3550 
3551 	/*
3552 	 * Re-lookup the vnode by path to detect a possible rename.
3553 	 * As a side effect, the vnode is relocked.
3554 	 * If vnode was renamed, return ENOENT.
3555 	 */
3556 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
3557 	    UIO_SYSSPACE, path, td);
3558 	error = namei(&nd);
3559 	if (error != 0) {
3560 		vrele(vp);
3561 		goto out;
3562 	}
3563 	NDFREE(&nd, NDF_ONLY_PNBUF);
3564 	vp1 = nd.ni_vp;
3565 	vrele(vp);
3566 	if (vp1 == vp)
3567 		strcpy(path, rpath);
3568 	else {
3569 		vput(vp1);
3570 		error = ENOENT;
3571 	}
3572 
3573 out:
3574 	free(fbuf, M_TEMP);
3575 	return (error);
3576 }
3577 
3578 #ifdef DDB
3579 static void
3580 db_print_vpath(struct vnode *vp)
3581 {
3582 
3583 	while (vp != NULL) {
3584 		db_printf("%p: ", vp);
3585 		if (vp == rootvnode) {
3586 			db_printf("/");
3587 			vp = NULL;
3588 		} else {
3589 			if (vp->v_vflag & VV_ROOT) {
3590 				db_printf("<mount point>");
3591 				vp = vp->v_mount->mnt_vnodecovered;
3592 			} else {
3593 				struct namecache *ncp;
3594 				char *ncn;
3595 				int i;
3596 
3597 				ncp = TAILQ_FIRST(&vp->v_cache_dst);
3598 				if (ncp != NULL) {
3599 					ncn = ncp->nc_name;
3600 					for (i = 0; i < ncp->nc_nlen; i++)
3601 						db_printf("%c", *ncn++);
3602 					vp = ncp->nc_dvp;
3603 				} else {
3604 					vp = NULL;
3605 				}
3606 			}
3607 		}
3608 		db_printf("\n");
3609 	}
3610 
3611 	return;
3612 }
3613 
3614 DB_SHOW_COMMAND(vpath, db_show_vpath)
3615 {
3616 	struct vnode *vp;
3617 
3618 	if (!have_addr) {
3619 		db_printf("usage: show vpath <struct vnode *>\n");
3620 		return;
3621 	}
3622 
3623 	vp = (struct vnode *)addr;
3624 	db_print_vpath(vp);
3625 }
3626 
3627 #endif
3628 
3629 static int cache_fast_lookup = 1;
3630 static char __read_frequently cache_fast_lookup_enabled = true;
3631 
3632 #define CACHE_FPL_FAILED	-2020
3633 
3634 void
3635 cache_fast_lookup_enabled_recalc(void)
3636 {
3637 	int lookup_flag;
3638 	int mac_on;
3639 
3640 #ifdef MAC
3641 	mac_on = mac_vnode_check_lookup_enabled();
3642 	mac_on |= mac_vnode_check_readlink_enabled();
3643 #else
3644 	mac_on = 0;
3645 #endif
3646 
3647 	lookup_flag = atomic_load_int(&cache_fast_lookup);
3648 	if (lookup_flag && !mac_on) {
3649 		atomic_store_char(&cache_fast_lookup_enabled, true);
3650 	} else {
3651 		atomic_store_char(&cache_fast_lookup_enabled, false);
3652 	}
3653 }
3654 
3655 static int
3656 syscal_vfs_cache_fast_lookup(SYSCTL_HANDLER_ARGS)
3657 {
3658 	int error, old;
3659 
3660 	old = atomic_load_int(&cache_fast_lookup);
3661 	error = sysctl_handle_int(oidp, arg1, arg2, req);
3662 	if (error == 0 && req->newptr && old != atomic_load_int(&cache_fast_lookup))
3663 		cache_fast_lookup_enabled_recalc();
3664 	return (error);
3665 }
3666 SYSCTL_PROC(_vfs, OID_AUTO, cache_fast_lookup, CTLTYPE_INT|CTLFLAG_RW|CTLFLAG_MPSAFE,
3667     &cache_fast_lookup, 0, syscal_vfs_cache_fast_lookup, "IU", "");
3668 
3669 /*
3670  * Components of nameidata (or objects it can point to) which may
3671  * need restoring in case fast path lookup fails.
3672  */
3673 struct nameidata_outer {
3674 	size_t ni_pathlen;
3675 	int cn_flags;
3676 };
3677 
3678 struct nameidata_saved {
3679 #ifdef INVARIANTS
3680 	char *cn_nameptr;
3681 	size_t ni_pathlen;
3682 #endif
3683 };
3684 
3685 #ifdef INVARIANTS
3686 struct cache_fpl_debug {
3687 	size_t ni_pathlen;
3688 };
3689 #endif
3690 
3691 struct cache_fpl {
3692 	struct nameidata *ndp;
3693 	struct componentname *cnp;
3694 	char *nulchar;
3695 	struct vnode *dvp;
3696 	struct vnode *tvp;
3697 	seqc_t dvp_seqc;
3698 	seqc_t tvp_seqc;
3699 	uint32_t hash;
3700 	struct nameidata_saved snd;
3701 	struct nameidata_outer snd_outer;
3702 	int line;
3703 	enum cache_fpl_status status:8;
3704 	bool in_smr;
3705 	bool fsearch;
3706 	bool savename;
3707 	struct pwd **pwd;
3708 #ifdef INVARIANTS
3709 	struct cache_fpl_debug debug;
3710 #endif
3711 };
3712 
3713 static bool cache_fplookup_is_mp(struct cache_fpl *fpl);
3714 static int cache_fplookup_cross_mount(struct cache_fpl *fpl);
3715 static int cache_fplookup_partial_setup(struct cache_fpl *fpl);
3716 static int cache_fplookup_skip_slashes(struct cache_fpl *fpl);
3717 static int cache_fplookup_trailingslash(struct cache_fpl *fpl);
3718 static void cache_fpl_pathlen_dec(struct cache_fpl *fpl);
3719 static void cache_fpl_pathlen_inc(struct cache_fpl *fpl);
3720 static void cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n);
3721 static void cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n);
3722 
3723 static void
3724 cache_fpl_cleanup_cnp(struct componentname *cnp)
3725 {
3726 
3727 	uma_zfree(namei_zone, cnp->cn_pnbuf);
3728 #ifdef DIAGNOSTIC
3729 	cnp->cn_pnbuf = NULL;
3730 	cnp->cn_nameptr = NULL;
3731 #endif
3732 }
3733 
3734 static struct vnode *
3735 cache_fpl_handle_root(struct cache_fpl *fpl)
3736 {
3737 	struct nameidata *ndp;
3738 	struct componentname *cnp;
3739 
3740 	ndp = fpl->ndp;
3741 	cnp = fpl->cnp;
3742 
3743 	MPASS(*(cnp->cn_nameptr) == '/');
3744 	cnp->cn_nameptr++;
3745 	cache_fpl_pathlen_dec(fpl);
3746 
3747 	if (__predict_false(*(cnp->cn_nameptr) == '/')) {
3748 		do {
3749 			cnp->cn_nameptr++;
3750 			cache_fpl_pathlen_dec(fpl);
3751 		} while (*(cnp->cn_nameptr) == '/');
3752 	}
3753 
3754 	return (ndp->ni_rootdir);
3755 }
3756 
3757 static void
3758 cache_fpl_checkpoint_outer(struct cache_fpl *fpl)
3759 {
3760 
3761 	fpl->snd_outer.ni_pathlen = fpl->ndp->ni_pathlen;
3762 	fpl->snd_outer.cn_flags = fpl->ndp->ni_cnd.cn_flags;
3763 }
3764 
3765 static void
3766 cache_fpl_checkpoint(struct cache_fpl *fpl)
3767 {
3768 
3769 #ifdef INVARIANTS
3770 	fpl->snd.cn_nameptr = fpl->ndp->ni_cnd.cn_nameptr;
3771 	fpl->snd.ni_pathlen = fpl->debug.ni_pathlen;
3772 #endif
3773 }
3774 
3775 static void
3776 cache_fpl_restore_partial(struct cache_fpl *fpl)
3777 {
3778 
3779 	fpl->ndp->ni_cnd.cn_flags = fpl->snd_outer.cn_flags;
3780 #ifdef INVARIANTS
3781 	fpl->debug.ni_pathlen = fpl->snd.ni_pathlen;
3782 #endif
3783 }
3784 
3785 static void
3786 cache_fpl_restore_abort(struct cache_fpl *fpl)
3787 {
3788 
3789 	cache_fpl_restore_partial(fpl);
3790 	/*
3791 	 * It is 0 on entry by API contract.
3792 	 */
3793 	fpl->ndp->ni_resflags = 0;
3794 	fpl->ndp->ni_cnd.cn_nameptr = fpl->ndp->ni_cnd.cn_pnbuf;
3795 	fpl->ndp->ni_pathlen = fpl->snd_outer.ni_pathlen;
3796 }
3797 
3798 #ifdef INVARIANTS
3799 #define cache_fpl_smr_assert_entered(fpl) ({			\
3800 	struct cache_fpl *_fpl = (fpl);				\
3801 	MPASS(_fpl->in_smr == true);				\
3802 	VFS_SMR_ASSERT_ENTERED();				\
3803 })
3804 #define cache_fpl_smr_assert_not_entered(fpl) ({		\
3805 	struct cache_fpl *_fpl = (fpl);				\
3806 	MPASS(_fpl->in_smr == false);				\
3807 	VFS_SMR_ASSERT_NOT_ENTERED();				\
3808 })
3809 static void
3810 cache_fpl_assert_status(struct cache_fpl *fpl)
3811 {
3812 
3813 	switch (fpl->status) {
3814 	case CACHE_FPL_STATUS_UNSET:
3815 		__assert_unreachable();
3816 		break;
3817 	case CACHE_FPL_STATUS_DESTROYED:
3818 	case CACHE_FPL_STATUS_ABORTED:
3819 	case CACHE_FPL_STATUS_PARTIAL:
3820 	case CACHE_FPL_STATUS_HANDLED:
3821 		break;
3822 	}
3823 }
3824 #else
3825 #define cache_fpl_smr_assert_entered(fpl) do { } while (0)
3826 #define cache_fpl_smr_assert_not_entered(fpl) do { } while (0)
3827 #define cache_fpl_assert_status(fpl) do { } while (0)
3828 #endif
3829 
3830 #define cache_fpl_smr_enter_initial(fpl) ({			\
3831 	struct cache_fpl *_fpl = (fpl);				\
3832 	vfs_smr_enter();					\
3833 	_fpl->in_smr = true;					\
3834 })
3835 
3836 #define cache_fpl_smr_enter(fpl) ({				\
3837 	struct cache_fpl *_fpl = (fpl);				\
3838 	MPASS(_fpl->in_smr == false);				\
3839 	vfs_smr_enter();					\
3840 	_fpl->in_smr = true;					\
3841 })
3842 
3843 #define cache_fpl_smr_exit(fpl) ({				\
3844 	struct cache_fpl *_fpl = (fpl);				\
3845 	MPASS(_fpl->in_smr == true);				\
3846 	vfs_smr_exit();						\
3847 	_fpl->in_smr = false;					\
3848 })
3849 
3850 static int
3851 cache_fpl_aborted_early_impl(struct cache_fpl *fpl, int line)
3852 {
3853 
3854 	if (fpl->status != CACHE_FPL_STATUS_UNSET) {
3855 		KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL,
3856 		    ("%s: converting to abort from %d at %d, set at %d\n",
3857 		    __func__, fpl->status, line, fpl->line));
3858 	}
3859 	cache_fpl_smr_assert_not_entered(fpl);
3860 	fpl->status = CACHE_FPL_STATUS_ABORTED;
3861 	fpl->line = line;
3862 	return (CACHE_FPL_FAILED);
3863 }
3864 
3865 #define cache_fpl_aborted_early(x)	cache_fpl_aborted_early_impl((x), __LINE__)
3866 
3867 static int __noinline
3868 cache_fpl_aborted_impl(struct cache_fpl *fpl, int line)
3869 {
3870 	struct nameidata *ndp;
3871 	struct componentname *cnp;
3872 
3873 	ndp = fpl->ndp;
3874 	cnp = fpl->cnp;
3875 
3876 	if (fpl->status != CACHE_FPL_STATUS_UNSET) {
3877 		KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL,
3878 		    ("%s: converting to abort from %d at %d, set at %d\n",
3879 		    __func__, fpl->status, line, fpl->line));
3880 	}
3881 	fpl->status = CACHE_FPL_STATUS_ABORTED;
3882 	fpl->line = line;
3883 	if (fpl->in_smr)
3884 		cache_fpl_smr_exit(fpl);
3885 	cache_fpl_restore_abort(fpl);
3886 	/*
3887 	 * Resolving symlinks overwrites data passed by the caller.
3888 	 * Let namei know.
3889 	 */
3890 	if (ndp->ni_loopcnt > 0) {
3891 		fpl->status = CACHE_FPL_STATUS_DESTROYED;
3892 		cache_fpl_cleanup_cnp(cnp);
3893 	}
3894 	return (CACHE_FPL_FAILED);
3895 }
3896 
3897 #define cache_fpl_aborted(x)	cache_fpl_aborted_impl((x), __LINE__)
3898 
3899 static int __noinline
3900 cache_fpl_partial_impl(struct cache_fpl *fpl, int line)
3901 {
3902 
3903 	KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3904 	    ("%s: setting to partial at %d, but already set to %d at %d\n",
3905 	    __func__, line, fpl->status, fpl->line));
3906 	cache_fpl_smr_assert_entered(fpl);
3907 	fpl->status = CACHE_FPL_STATUS_PARTIAL;
3908 	fpl->line = line;
3909 	return (cache_fplookup_partial_setup(fpl));
3910 }
3911 
3912 #define cache_fpl_partial(x)	cache_fpl_partial_impl((x), __LINE__)
3913 
3914 static int
3915 cache_fpl_handled_impl(struct cache_fpl *fpl, int line)
3916 {
3917 
3918 	KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3919 	    ("%s: setting to handled at %d, but already set to %d at %d\n",
3920 	    __func__, line, fpl->status, fpl->line));
3921 	cache_fpl_smr_assert_not_entered(fpl);
3922 	fpl->status = CACHE_FPL_STATUS_HANDLED;
3923 	fpl->line = line;
3924 	return (0);
3925 }
3926 
3927 #define cache_fpl_handled(x)	cache_fpl_handled_impl((x), __LINE__)
3928 
3929 static int
3930 cache_fpl_handled_error_impl(struct cache_fpl *fpl, int error, int line)
3931 {
3932 
3933 	KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3934 	    ("%s: setting to handled at %d, but already set to %d at %d\n",
3935 	    __func__, line, fpl->status, fpl->line));
3936 	MPASS(error != 0);
3937 	MPASS(error != CACHE_FPL_FAILED);
3938 	cache_fpl_smr_assert_not_entered(fpl);
3939 	fpl->status = CACHE_FPL_STATUS_HANDLED;
3940 	fpl->line = line;
3941 	fpl->dvp = NULL;
3942 	fpl->tvp = NULL;
3943 	fpl->savename = false;
3944 	return (error);
3945 }
3946 
3947 #define cache_fpl_handled_error(x, e)	cache_fpl_handled_error_impl((x), (e), __LINE__)
3948 
3949 static bool
3950 cache_fpl_terminated(struct cache_fpl *fpl)
3951 {
3952 
3953 	return (fpl->status != CACHE_FPL_STATUS_UNSET);
3954 }
3955 
3956 #define CACHE_FPL_SUPPORTED_CN_FLAGS \
3957 	(NC_NOMAKEENTRY | NC_KEEPPOSENTRY | LOCKLEAF | LOCKPARENT | WANTPARENT | \
3958 	 FAILIFEXISTS | FOLLOW | LOCKSHARED | SAVENAME | SAVESTART | WILLBEDIR | \
3959 	 ISOPEN | NOMACCHECK | AUDITVNODE1 | AUDITVNODE2 | NOCAPCHECK)
3960 
3961 #define CACHE_FPL_INTERNAL_CN_FLAGS \
3962 	(ISDOTDOT | MAKEENTRY | ISLASTCN)
3963 
3964 _Static_assert((CACHE_FPL_SUPPORTED_CN_FLAGS & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
3965     "supported and internal flags overlap");
3966 
3967 static bool
3968 cache_fpl_islastcn(struct nameidata *ndp)
3969 {
3970 
3971 	return (*ndp->ni_next == 0);
3972 }
3973 
3974 static bool
3975 cache_fpl_istrailingslash(struct cache_fpl *fpl)
3976 {
3977 
3978 	return (*(fpl->nulchar - 1) == '/');
3979 }
3980 
3981 static bool
3982 cache_fpl_isdotdot(struct componentname *cnp)
3983 {
3984 
3985 	if (cnp->cn_namelen == 2 &&
3986 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
3987 		return (true);
3988 	return (false);
3989 }
3990 
3991 static bool
3992 cache_can_fplookup(struct cache_fpl *fpl)
3993 {
3994 	struct nameidata *ndp;
3995 	struct componentname *cnp;
3996 	struct thread *td;
3997 
3998 	ndp = fpl->ndp;
3999 	cnp = fpl->cnp;
4000 	td = cnp->cn_thread;
4001 
4002 	if (!atomic_load_char(&cache_fast_lookup_enabled)) {
4003 		cache_fpl_aborted_early(fpl);
4004 		return (false);
4005 	}
4006 	if ((cnp->cn_flags & ~CACHE_FPL_SUPPORTED_CN_FLAGS) != 0) {
4007 		cache_fpl_aborted_early(fpl);
4008 		return (false);
4009 	}
4010 	if (IN_CAPABILITY_MODE(td)) {
4011 		cache_fpl_aborted_early(fpl);
4012 		return (false);
4013 	}
4014 	if (AUDITING_TD(td)) {
4015 		cache_fpl_aborted_early(fpl);
4016 		return (false);
4017 	}
4018 	if (ndp->ni_startdir != NULL) {
4019 		cache_fpl_aborted_early(fpl);
4020 		return (false);
4021 	}
4022 	return (true);
4023 }
4024 
4025 static int
4026 cache_fplookup_dirfd(struct cache_fpl *fpl, struct vnode **vpp)
4027 {
4028 	struct nameidata *ndp;
4029 	int error;
4030 	bool fsearch;
4031 
4032 	ndp = fpl->ndp;
4033 	error = fgetvp_lookup_smr(ndp->ni_dirfd, ndp, vpp, &fsearch);
4034 	if (__predict_false(error != 0)) {
4035 		return (cache_fpl_aborted(fpl));
4036 	}
4037 	fpl->fsearch = fsearch;
4038 	return (0);
4039 }
4040 
4041 static int __noinline
4042 cache_fplookup_negative_promote(struct cache_fpl *fpl, struct namecache *oncp,
4043     uint32_t hash)
4044 {
4045 	struct componentname *cnp;
4046 	struct vnode *dvp;
4047 
4048 	cnp = fpl->cnp;
4049 	dvp = fpl->dvp;
4050 
4051 	cache_fpl_smr_exit(fpl);
4052 	if (cache_neg_promote_cond(dvp, cnp, oncp, hash))
4053 		return (cache_fpl_handled_error(fpl, ENOENT));
4054 	else
4055 		return (cache_fpl_aborted(fpl));
4056 }
4057 
4058 /*
4059  * The target vnode is not supported, prepare for the slow path to take over.
4060  */
4061 static int __noinline
4062 cache_fplookup_partial_setup(struct cache_fpl *fpl)
4063 {
4064 	struct nameidata *ndp;
4065 	struct componentname *cnp;
4066 	enum vgetstate dvs;
4067 	struct vnode *dvp;
4068 	struct pwd *pwd;
4069 	seqc_t dvp_seqc;
4070 
4071 	ndp = fpl->ndp;
4072 	cnp = fpl->cnp;
4073 	pwd = *(fpl->pwd);
4074 	dvp = fpl->dvp;
4075 	dvp_seqc = fpl->dvp_seqc;
4076 
4077 	if (!pwd_hold_smr(pwd)) {
4078 		return (cache_fpl_aborted(fpl));
4079 	}
4080 
4081 	/*
4082 	 * Note that seqc is checked before the vnode is locked, so by
4083 	 * the time regular lookup gets to it it may have moved.
4084 	 *
4085 	 * Ultimately this does not affect correctness, any lookup errors
4086 	 * are userspace racing with itself. It is guaranteed that any
4087 	 * path which ultimately gets found could also have been found
4088 	 * by regular lookup going all the way in absence of concurrent
4089 	 * modifications.
4090 	 */
4091 	dvs = vget_prep_smr(dvp);
4092 	cache_fpl_smr_exit(fpl);
4093 	if (__predict_false(dvs == VGET_NONE)) {
4094 		pwd_drop(pwd);
4095 		return (cache_fpl_aborted(fpl));
4096 	}
4097 
4098 	vget_finish_ref(dvp, dvs);
4099 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4100 		vrele(dvp);
4101 		pwd_drop(pwd);
4102 		return (cache_fpl_aborted(fpl));
4103 	}
4104 
4105 	cache_fpl_restore_partial(fpl);
4106 #ifdef INVARIANTS
4107 	if (cnp->cn_nameptr != fpl->snd.cn_nameptr) {
4108 		panic("%s: cn_nameptr mismatch (%p != %p) full [%s]\n", __func__,
4109 		    cnp->cn_nameptr, fpl->snd.cn_nameptr, cnp->cn_pnbuf);
4110 	}
4111 #endif
4112 
4113 	ndp->ni_startdir = dvp;
4114 	cnp->cn_flags |= MAKEENTRY;
4115 	if (cache_fpl_islastcn(ndp))
4116 		cnp->cn_flags |= ISLASTCN;
4117 	if (cache_fpl_isdotdot(cnp))
4118 		cnp->cn_flags |= ISDOTDOT;
4119 
4120 	/*
4121 	 * Skip potential extra slashes parsing did not take care of.
4122 	 * cache_fplookup_skip_slashes explains the mechanism.
4123 	 */
4124 	if (__predict_false(*(cnp->cn_nameptr) == '/')) {
4125 		do {
4126 			cnp->cn_nameptr++;
4127 			cache_fpl_pathlen_dec(fpl);
4128 		} while (*(cnp->cn_nameptr) == '/');
4129 	}
4130 
4131 	ndp->ni_pathlen = fpl->nulchar - cnp->cn_nameptr + 1;
4132 #ifdef INVARIANTS
4133 	if (ndp->ni_pathlen != fpl->debug.ni_pathlen) {
4134 		panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n",
4135 		    __func__, ndp->ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar,
4136 		    cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf);
4137 	}
4138 #endif
4139 	return (0);
4140 }
4141 
4142 static int
4143 cache_fplookup_final_child(struct cache_fpl *fpl, enum vgetstate tvs)
4144 {
4145 	struct componentname *cnp;
4146 	struct vnode *tvp;
4147 	seqc_t tvp_seqc;
4148 	int error, lkflags;
4149 
4150 	cnp = fpl->cnp;
4151 	tvp = fpl->tvp;
4152 	tvp_seqc = fpl->tvp_seqc;
4153 
4154 	if ((cnp->cn_flags & LOCKLEAF) != 0) {
4155 		lkflags = LK_SHARED;
4156 		if ((cnp->cn_flags & LOCKSHARED) == 0)
4157 			lkflags = LK_EXCLUSIVE;
4158 		error = vget_finish(tvp, lkflags, tvs);
4159 		if (__predict_false(error != 0)) {
4160 			return (cache_fpl_aborted(fpl));
4161 		}
4162 	} else {
4163 		vget_finish_ref(tvp, tvs);
4164 	}
4165 
4166 	if (!vn_seqc_consistent(tvp, tvp_seqc)) {
4167 		if ((cnp->cn_flags & LOCKLEAF) != 0)
4168 			vput(tvp);
4169 		else
4170 			vrele(tvp);
4171 		return (cache_fpl_aborted(fpl));
4172 	}
4173 
4174 	return (cache_fpl_handled(fpl));
4175 }
4176 
4177 /*
4178  * They want to possibly modify the state of the namecache.
4179  */
4180 static int __noinline
4181 cache_fplookup_final_modifying(struct cache_fpl *fpl)
4182 {
4183 	struct nameidata *ndp;
4184 	struct componentname *cnp;
4185 	enum vgetstate dvs;
4186 	struct vnode *dvp, *tvp;
4187 	struct mount *mp;
4188 	seqc_t dvp_seqc;
4189 	int error;
4190 	bool docache;
4191 
4192 	ndp = fpl->ndp;
4193 	cnp = fpl->cnp;
4194 	dvp = fpl->dvp;
4195 	dvp_seqc = fpl->dvp_seqc;
4196 
4197 	MPASS(*(cnp->cn_nameptr) != '/');
4198 	MPASS(cache_fpl_islastcn(ndp));
4199 	if ((cnp->cn_flags & LOCKPARENT) == 0)
4200 		MPASS((cnp->cn_flags & WANTPARENT) != 0);
4201 	MPASS((cnp->cn_flags & TRAILINGSLASH) == 0);
4202 	MPASS(cnp->cn_nameiop == CREATE || cnp->cn_nameiop == DELETE ||
4203 	    cnp->cn_nameiop == RENAME);
4204 	MPASS((cnp->cn_flags & MAKEENTRY) == 0);
4205 	MPASS((cnp->cn_flags & ISDOTDOT) == 0);
4206 
4207 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
4208 	if (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)
4209 		docache = false;
4210 
4211 	/*
4212 	 * Regular lookup nulifies the slash, which we don't do here.
4213 	 * Don't take chances with filesystem routines seeing it for
4214 	 * the last entry.
4215 	 */
4216 	if (cache_fpl_istrailingslash(fpl)) {
4217 		return (cache_fpl_partial(fpl));
4218 	}
4219 
4220 	mp = atomic_load_ptr(&dvp->v_mount);
4221 	if (__predict_false(mp == NULL)) {
4222 		return (cache_fpl_aborted(fpl));
4223 	}
4224 
4225 	if (__predict_false(mp->mnt_flag & MNT_RDONLY)) {
4226 		cache_fpl_smr_exit(fpl);
4227 		/*
4228 		 * Original code keeps not checking for CREATE which
4229 		 * might be a bug. For now let the old lookup decide.
4230 		 */
4231 		if (cnp->cn_nameiop == CREATE) {
4232 			return (cache_fpl_aborted(fpl));
4233 		}
4234 		return (cache_fpl_handled_error(fpl, EROFS));
4235 	}
4236 
4237 	if (fpl->tvp != NULL && (cnp->cn_flags & FAILIFEXISTS) != 0) {
4238 		cache_fpl_smr_exit(fpl);
4239 		return (cache_fpl_handled_error(fpl, EEXIST));
4240 	}
4241 
4242 	/*
4243 	 * Secure access to dvp; check cache_fplookup_partial_setup for
4244 	 * reasoning.
4245 	 *
4246 	 * XXX At least UFS requires its lookup routine to be called for
4247 	 * the last path component, which leads to some level of complication
4248 	 * and inefficiency:
4249 	 * - the target routine always locks the target vnode, but our caller
4250 	 *   may not need it locked
4251 	 * - some of the VOP machinery asserts that the parent is locked, which
4252 	 *   once more may be not required
4253 	 *
4254 	 * TODO: add a flag for filesystems which don't need this.
4255 	 */
4256 	dvs = vget_prep_smr(dvp);
4257 	cache_fpl_smr_exit(fpl);
4258 	if (__predict_false(dvs == VGET_NONE)) {
4259 		return (cache_fpl_aborted(fpl));
4260 	}
4261 
4262 	vget_finish_ref(dvp, dvs);
4263 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4264 		vrele(dvp);
4265 		return (cache_fpl_aborted(fpl));
4266 	}
4267 
4268 	error = vn_lock(dvp, LK_EXCLUSIVE);
4269 	if (__predict_false(error != 0)) {
4270 		vrele(dvp);
4271 		return (cache_fpl_aborted(fpl));
4272 	}
4273 
4274 	tvp = NULL;
4275 	cnp->cn_flags |= ISLASTCN;
4276 	if (docache)
4277 		cnp->cn_flags |= MAKEENTRY;
4278 	if (cache_fpl_isdotdot(cnp))
4279 		cnp->cn_flags |= ISDOTDOT;
4280 	cnp->cn_lkflags = LK_EXCLUSIVE;
4281 	error = VOP_LOOKUP(dvp, &tvp, cnp);
4282 	switch (error) {
4283 	case EJUSTRETURN:
4284 	case 0:
4285 		break;
4286 	case ENOTDIR:
4287 	case ENOENT:
4288 		vput(dvp);
4289 		return (cache_fpl_handled_error(fpl, error));
4290 	default:
4291 		vput(dvp);
4292 		return (cache_fpl_aborted(fpl));
4293 	}
4294 
4295 	fpl->tvp = tvp;
4296 	fpl->savename = (cnp->cn_flags & SAVENAME) != 0;
4297 
4298 	if (tvp == NULL) {
4299 		if ((cnp->cn_flags & SAVESTART) != 0) {
4300 			ndp->ni_startdir = dvp;
4301 			vrefact(ndp->ni_startdir);
4302 			cnp->cn_flags |= SAVENAME;
4303 			fpl->savename = true;
4304 		}
4305 		MPASS(error == EJUSTRETURN);
4306 		if ((cnp->cn_flags & LOCKPARENT) == 0) {
4307 			VOP_UNLOCK(dvp);
4308 		}
4309 		return (cache_fpl_handled(fpl));
4310 	}
4311 
4312 	/*
4313 	 * There are very hairy corner cases concerning various flag combinations
4314 	 * and locking state. In particular here we only hold one lock instead of
4315 	 * two.
4316 	 *
4317 	 * Skip the complexity as it is of no significance for normal workloads.
4318 	 */
4319 	if (__predict_false(tvp == dvp)) {
4320 		vput(dvp);
4321 		vrele(tvp);
4322 		return (cache_fpl_aborted(fpl));
4323 	}
4324 
4325 	/*
4326 	 * If they want the symlink itself we are fine, but if they want to
4327 	 * follow it regular lookup has to be engaged.
4328 	 */
4329 	if (tvp->v_type == VLNK) {
4330 		if ((cnp->cn_flags & FOLLOW) != 0) {
4331 			vput(dvp);
4332 			vput(tvp);
4333 			return (cache_fpl_aborted(fpl));
4334 		}
4335 	}
4336 
4337 	/*
4338 	 * Since we expect this to be the terminal vnode it should almost never
4339 	 * be a mount point.
4340 	 */
4341 	if (__predict_false(cache_fplookup_is_mp(fpl))) {
4342 		vput(dvp);
4343 		vput(tvp);
4344 		return (cache_fpl_aborted(fpl));
4345 	}
4346 
4347 	if ((cnp->cn_flags & FAILIFEXISTS) != 0) {
4348 		vput(dvp);
4349 		vput(tvp);
4350 		return (cache_fpl_handled_error(fpl, EEXIST));
4351 	}
4352 
4353 	if ((cnp->cn_flags & LOCKLEAF) == 0) {
4354 		VOP_UNLOCK(tvp);
4355 	}
4356 
4357 	if ((cnp->cn_flags & LOCKPARENT) == 0) {
4358 		VOP_UNLOCK(dvp);
4359 	}
4360 
4361 	if ((cnp->cn_flags & SAVESTART) != 0) {
4362 		ndp->ni_startdir = dvp;
4363 		vrefact(ndp->ni_startdir);
4364 		cnp->cn_flags |= SAVENAME;
4365 		fpl->savename = true;
4366 	}
4367 
4368 	return (cache_fpl_handled(fpl));
4369 }
4370 
4371 static int __noinline
4372 cache_fplookup_modifying(struct cache_fpl *fpl)
4373 {
4374 	struct nameidata *ndp;
4375 
4376 	ndp = fpl->ndp;
4377 
4378 	if (!cache_fpl_islastcn(ndp)) {
4379 		return (cache_fpl_partial(fpl));
4380 	}
4381 	return (cache_fplookup_final_modifying(fpl));
4382 }
4383 
4384 static int __noinline
4385 cache_fplookup_final_withparent(struct cache_fpl *fpl)
4386 {
4387 	struct componentname *cnp;
4388 	enum vgetstate dvs, tvs;
4389 	struct vnode *dvp, *tvp;
4390 	seqc_t dvp_seqc;
4391 	int error;
4392 
4393 	cnp = fpl->cnp;
4394 	dvp = fpl->dvp;
4395 	dvp_seqc = fpl->dvp_seqc;
4396 	tvp = fpl->tvp;
4397 
4398 	MPASS((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0);
4399 
4400 	/*
4401 	 * This is less efficient than it can be for simplicity.
4402 	 */
4403 	dvs = vget_prep_smr(dvp);
4404 	if (__predict_false(dvs == VGET_NONE)) {
4405 		return (cache_fpl_aborted(fpl));
4406 	}
4407 	tvs = vget_prep_smr(tvp);
4408 	if (__predict_false(tvs == VGET_NONE)) {
4409 		cache_fpl_smr_exit(fpl);
4410 		vget_abort(dvp, dvs);
4411 		return (cache_fpl_aborted(fpl));
4412 	}
4413 
4414 	cache_fpl_smr_exit(fpl);
4415 
4416 	if ((cnp->cn_flags & LOCKPARENT) != 0) {
4417 		error = vget_finish(dvp, LK_EXCLUSIVE, dvs);
4418 		if (__predict_false(error != 0)) {
4419 			vget_abort(tvp, tvs);
4420 			return (cache_fpl_aborted(fpl));
4421 		}
4422 	} else {
4423 		vget_finish_ref(dvp, dvs);
4424 	}
4425 
4426 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4427 		vget_abort(tvp, tvs);
4428 		if ((cnp->cn_flags & LOCKPARENT) != 0)
4429 			vput(dvp);
4430 		else
4431 			vrele(dvp);
4432 		return (cache_fpl_aborted(fpl));
4433 	}
4434 
4435 	error = cache_fplookup_final_child(fpl, tvs);
4436 	if (__predict_false(error != 0)) {
4437 		MPASS(fpl->status == CACHE_FPL_STATUS_ABORTED);
4438 		if ((cnp->cn_flags & LOCKPARENT) != 0)
4439 			vput(dvp);
4440 		else
4441 			vrele(dvp);
4442 		return (error);
4443 	}
4444 
4445 	MPASS(fpl->status == CACHE_FPL_STATUS_HANDLED);
4446 	return (0);
4447 }
4448 
4449 static int
4450 cache_fplookup_final(struct cache_fpl *fpl)
4451 {
4452 	struct componentname *cnp;
4453 	enum vgetstate tvs;
4454 	struct vnode *dvp, *tvp;
4455 	seqc_t dvp_seqc;
4456 
4457 	cnp = fpl->cnp;
4458 	dvp = fpl->dvp;
4459 	dvp_seqc = fpl->dvp_seqc;
4460 	tvp = fpl->tvp;
4461 
4462 	MPASS(*(cnp->cn_nameptr) != '/');
4463 
4464 	if (cnp->cn_nameiop != LOOKUP) {
4465 		return (cache_fplookup_final_modifying(fpl));
4466 	}
4467 
4468 	if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0)
4469 		return (cache_fplookup_final_withparent(fpl));
4470 
4471 	tvs = vget_prep_smr(tvp);
4472 	if (__predict_false(tvs == VGET_NONE)) {
4473 		return (cache_fpl_partial(fpl));
4474 	}
4475 
4476 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4477 		cache_fpl_smr_exit(fpl);
4478 		vget_abort(tvp, tvs);
4479 		return (cache_fpl_aborted(fpl));
4480 	}
4481 
4482 	cache_fpl_smr_exit(fpl);
4483 	return (cache_fplookup_final_child(fpl, tvs));
4484 }
4485 
4486 /*
4487  * Comment from locked lookup:
4488  * Check for degenerate name (e.g. / or "") which is a way of talking about a
4489  * directory, e.g. like "/." or ".".
4490  */
4491 static int __noinline
4492 cache_fplookup_degenerate(struct cache_fpl *fpl)
4493 {
4494 	struct componentname *cnp;
4495 	struct vnode *dvp;
4496 	enum vgetstate dvs;
4497 	int error, lkflags;
4498 #ifdef INVARIANTS
4499 	char *cp;
4500 #endif
4501 
4502 	fpl->tvp = fpl->dvp;
4503 	fpl->tvp_seqc = fpl->dvp_seqc;
4504 
4505 	cnp = fpl->cnp;
4506 	dvp = fpl->dvp;
4507 
4508 #ifdef INVARIANTS
4509 	for (cp = cnp->cn_pnbuf; *cp != '\0'; cp++) {
4510 		KASSERT(*cp == '/',
4511 		    ("%s: encountered non-slash; string [%s]\n", __func__,
4512 		    cnp->cn_pnbuf));
4513 	}
4514 #endif
4515 
4516 	if (__predict_false(cnp->cn_nameiop != LOOKUP)) {
4517 		cache_fpl_smr_exit(fpl);
4518 		return (cache_fpl_handled_error(fpl, EISDIR));
4519 	}
4520 
4521 	MPASS((cnp->cn_flags & SAVESTART) == 0);
4522 
4523 	if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0) {
4524 		return (cache_fplookup_final_withparent(fpl));
4525 	}
4526 
4527 	dvs = vget_prep_smr(dvp);
4528 	cache_fpl_smr_exit(fpl);
4529 	if (__predict_false(dvs == VGET_NONE)) {
4530 		return (cache_fpl_aborted(fpl));
4531 	}
4532 
4533 	if ((cnp->cn_flags & LOCKLEAF) != 0) {
4534 		lkflags = LK_SHARED;
4535 		if ((cnp->cn_flags & LOCKSHARED) == 0)
4536 			lkflags = LK_EXCLUSIVE;
4537 		error = vget_finish(dvp, lkflags, dvs);
4538 		if (__predict_false(error != 0)) {
4539 			return (cache_fpl_aborted(fpl));
4540 		}
4541 	} else {
4542 		vget_finish_ref(dvp, dvs);
4543 	}
4544 	return (cache_fpl_handled(fpl));
4545 }
4546 
4547 static int __noinline
4548 cache_fplookup_noentry(struct cache_fpl *fpl)
4549 {
4550 	struct nameidata *ndp;
4551 	struct componentname *cnp;
4552 	enum vgetstate dvs;
4553 	struct vnode *dvp, *tvp;
4554 	seqc_t dvp_seqc;
4555 	int error;
4556 	bool docache;
4557 
4558 	ndp = fpl->ndp;
4559 	cnp = fpl->cnp;
4560 	dvp = fpl->dvp;
4561 	dvp_seqc = fpl->dvp_seqc;
4562 
4563 	MPASS((cnp->cn_flags & MAKEENTRY) == 0);
4564 	MPASS((cnp->cn_flags & ISDOTDOT) == 0);
4565 	MPASS(!cache_fpl_isdotdot(cnp));
4566 
4567 	/*
4568 	 * Hack: delayed name len checking.
4569 	 */
4570 	if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
4571 		cache_fpl_smr_exit(fpl);
4572 		return (cache_fpl_handled_error(fpl, ENAMETOOLONG));
4573 	}
4574 
4575 	if (cnp->cn_nameptr[0] == '/') {
4576 		return (cache_fplookup_skip_slashes(fpl));
4577 	}
4578 
4579 	if (cnp->cn_nameptr[0] == '\0') {
4580 		if (fpl->tvp == NULL) {
4581 			return (cache_fplookup_degenerate(fpl));
4582 		}
4583 		return (cache_fplookup_trailingslash(fpl));
4584 	}
4585 
4586 	if (cnp->cn_nameiop != LOOKUP) {
4587 		fpl->tvp = NULL;
4588 		return (cache_fplookup_modifying(fpl));
4589 	}
4590 
4591 	MPASS((cnp->cn_flags & SAVESTART) == 0);
4592 
4593 	/*
4594 	 * Only try to fill in the component if it is the last one,
4595 	 * otherwise not only there may be several to handle but the
4596 	 * walk may be complicated.
4597 	 */
4598 	if (!cache_fpl_islastcn(ndp)) {
4599 		return (cache_fpl_partial(fpl));
4600 	}
4601 
4602 	/*
4603 	 * Regular lookup nulifies the slash, which we don't do here.
4604 	 * Don't take chances with filesystem routines seeing it for
4605 	 * the last entry.
4606 	 */
4607 	if (cache_fpl_istrailingslash(fpl)) {
4608 		return (cache_fpl_partial(fpl));
4609 	}
4610 
4611 	/*
4612 	 * Secure access to dvp; check cache_fplookup_partial_setup for
4613 	 * reasoning.
4614 	 */
4615 	dvs = vget_prep_smr(dvp);
4616 	cache_fpl_smr_exit(fpl);
4617 	if (__predict_false(dvs == VGET_NONE)) {
4618 		return (cache_fpl_aborted(fpl));
4619 	}
4620 
4621 	vget_finish_ref(dvp, dvs);
4622 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4623 		vrele(dvp);
4624 		return (cache_fpl_aborted(fpl));
4625 	}
4626 
4627 	error = vn_lock(dvp, LK_SHARED);
4628 	if (__predict_false(error != 0)) {
4629 		vrele(dvp);
4630 		return (cache_fpl_aborted(fpl));
4631 	}
4632 
4633 	tvp = NULL;
4634 	/*
4635 	 * TODO: provide variants which don't require locking either vnode.
4636 	 */
4637 	cnp->cn_flags |= ISLASTCN;
4638 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
4639 	if (docache)
4640 		cnp->cn_flags |= MAKEENTRY;
4641 	cnp->cn_lkflags = LK_SHARED;
4642 	if ((cnp->cn_flags & LOCKSHARED) == 0) {
4643 		cnp->cn_lkflags = LK_EXCLUSIVE;
4644 	}
4645 	error = VOP_LOOKUP(dvp, &tvp, cnp);
4646 	switch (error) {
4647 	case EJUSTRETURN:
4648 	case 0:
4649 		break;
4650 	case ENOTDIR:
4651 	case ENOENT:
4652 		vput(dvp);
4653 		return (cache_fpl_handled_error(fpl, error));
4654 	default:
4655 		vput(dvp);
4656 		return (cache_fpl_aborted(fpl));
4657 	}
4658 
4659 	fpl->tvp = tvp;
4660 	if (!fpl->savename) {
4661 		MPASS((cnp->cn_flags & SAVENAME) == 0);
4662 	}
4663 
4664 	if (tvp == NULL) {
4665 		MPASS(error == EJUSTRETURN);
4666 		if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) {
4667 			vput(dvp);
4668 		} else if ((cnp->cn_flags & LOCKPARENT) == 0) {
4669 			VOP_UNLOCK(dvp);
4670 		}
4671 		return (cache_fpl_handled(fpl));
4672 	}
4673 
4674 	if (tvp->v_type == VLNK) {
4675 		if ((cnp->cn_flags & FOLLOW) != 0) {
4676 			vput(dvp);
4677 			vput(tvp);
4678 			return (cache_fpl_aborted(fpl));
4679 		}
4680 	}
4681 
4682 	if (__predict_false(cache_fplookup_is_mp(fpl))) {
4683 		vput(dvp);
4684 		vput(tvp);
4685 		return (cache_fpl_aborted(fpl));
4686 	}
4687 
4688 	if ((cnp->cn_flags & LOCKLEAF) == 0) {
4689 		VOP_UNLOCK(tvp);
4690 	}
4691 
4692 	if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) {
4693 		vput(dvp);
4694 	} else if ((cnp->cn_flags & LOCKPARENT) == 0) {
4695 		VOP_UNLOCK(dvp);
4696 	}
4697 	return (cache_fpl_handled(fpl));
4698 }
4699 
4700 static int __noinline
4701 cache_fplookup_dot(struct cache_fpl *fpl)
4702 {
4703 	int error;
4704 
4705 	MPASS(!seqc_in_modify(fpl->dvp_seqc));
4706 	/*
4707 	 * Just re-assign the value. seqc will be checked later for the first
4708 	 * non-dot path component in line and/or before deciding to return the
4709 	 * vnode.
4710 	 */
4711 	fpl->tvp = fpl->dvp;
4712 	fpl->tvp_seqc = fpl->dvp_seqc;
4713 
4714 	counter_u64_add(dothits, 1);
4715 	SDT_PROBE3(vfs, namecache, lookup, hit, fpl->dvp, ".", fpl->dvp);
4716 
4717 	error = 0;
4718 	if (cache_fplookup_is_mp(fpl)) {
4719 		error = cache_fplookup_cross_mount(fpl);
4720 	}
4721 	return (error);
4722 }
4723 
4724 static int __noinline
4725 cache_fplookup_dotdot(struct cache_fpl *fpl)
4726 {
4727 	struct nameidata *ndp;
4728 	struct componentname *cnp;
4729 	struct namecache *ncp;
4730 	struct vnode *dvp;
4731 	struct prison *pr;
4732 	u_char nc_flag;
4733 
4734 	ndp = fpl->ndp;
4735 	cnp = fpl->cnp;
4736 	dvp = fpl->dvp;
4737 
4738 	MPASS(cache_fpl_isdotdot(cnp));
4739 
4740 	/*
4741 	 * XXX this is racy the same way regular lookup is
4742 	 */
4743 	for (pr = cnp->cn_cred->cr_prison; pr != NULL;
4744 	    pr = pr->pr_parent)
4745 		if (dvp == pr->pr_root)
4746 			break;
4747 
4748 	if (dvp == ndp->ni_rootdir ||
4749 	    dvp == ndp->ni_topdir ||
4750 	    dvp == rootvnode ||
4751 	    pr != NULL) {
4752 		fpl->tvp = dvp;
4753 		fpl->tvp_seqc = vn_seqc_read_any(dvp);
4754 		if (seqc_in_modify(fpl->tvp_seqc)) {
4755 			return (cache_fpl_aborted(fpl));
4756 		}
4757 		return (0);
4758 	}
4759 
4760 	if ((dvp->v_vflag & VV_ROOT) != 0) {
4761 		/*
4762 		 * TODO
4763 		 * The opposite of climb mount is needed here.
4764 		 */
4765 		return (cache_fpl_partial(fpl));
4766 	}
4767 
4768 	ncp = atomic_load_consume_ptr(&dvp->v_cache_dd);
4769 	if (ncp == NULL) {
4770 		return (cache_fpl_aborted(fpl));
4771 	}
4772 
4773 	nc_flag = atomic_load_char(&ncp->nc_flag);
4774 	if ((nc_flag & NCF_ISDOTDOT) != 0) {
4775 		if ((nc_flag & NCF_NEGATIVE) != 0)
4776 			return (cache_fpl_aborted(fpl));
4777 		fpl->tvp = ncp->nc_vp;
4778 	} else {
4779 		fpl->tvp = ncp->nc_dvp;
4780 	}
4781 
4782 	fpl->tvp_seqc = vn_seqc_read_any(fpl->tvp);
4783 	if (seqc_in_modify(fpl->tvp_seqc)) {
4784 		return (cache_fpl_partial(fpl));
4785 	}
4786 
4787 	/*
4788 	 * Acquire fence provided by vn_seqc_read_any above.
4789 	 */
4790 	if (__predict_false(atomic_load_ptr(&dvp->v_cache_dd) != ncp)) {
4791 		return (cache_fpl_aborted(fpl));
4792 	}
4793 
4794 	if (!cache_ncp_canuse(ncp)) {
4795 		return (cache_fpl_aborted(fpl));
4796 	}
4797 
4798 	counter_u64_add(dotdothits, 1);
4799 	return (0);
4800 }
4801 
4802 static int __noinline
4803 cache_fplookup_neg(struct cache_fpl *fpl, struct namecache *ncp, uint32_t hash)
4804 {
4805 	u_char nc_flag;
4806 	bool neg_promote;
4807 
4808 	nc_flag = atomic_load_char(&ncp->nc_flag);
4809 	MPASS((nc_flag & NCF_NEGATIVE) != 0);
4810 	/*
4811 	 * If they want to create an entry we need to replace this one.
4812 	 */
4813 	if (__predict_false(fpl->cnp->cn_nameiop != LOOKUP)) {
4814 		fpl->tvp = NULL;
4815 		return (cache_fplookup_modifying(fpl));
4816 	}
4817 	neg_promote = cache_neg_hit_prep(ncp);
4818 	if (!cache_fpl_neg_ncp_canuse(ncp)) {
4819 		cache_neg_hit_abort(ncp);
4820 		return (cache_fpl_partial(fpl));
4821 	}
4822 	if (neg_promote) {
4823 		return (cache_fplookup_negative_promote(fpl, ncp, hash));
4824 	}
4825 	cache_neg_hit_finish(ncp);
4826 	cache_fpl_smr_exit(fpl);
4827 	return (cache_fpl_handled_error(fpl, ENOENT));
4828 }
4829 
4830 /*
4831  * Resolve a symlink. Called by filesystem-specific routines.
4832  *
4833  * Code flow is:
4834  * ... -> cache_fplookup_symlink -> VOP_FPLOOKUP_SYMLINK -> cache_symlink_resolve
4835  */
4836 int
4837 cache_symlink_resolve(struct cache_fpl *fpl, const char *string, size_t len)
4838 {
4839 	struct nameidata *ndp;
4840 	struct componentname *cnp;
4841 	size_t adjust;
4842 
4843 	ndp = fpl->ndp;
4844 	cnp = fpl->cnp;
4845 
4846 	if (__predict_false(len == 0)) {
4847 		return (ENOENT);
4848 	}
4849 
4850 	if (__predict_false(len > MAXPATHLEN - 2)) {
4851 		if (cache_fpl_istrailingslash(fpl)) {
4852 			return (EAGAIN);
4853 		}
4854 	}
4855 
4856 	ndp->ni_pathlen = fpl->nulchar - cnp->cn_nameptr - cnp->cn_namelen + 1;
4857 #ifdef INVARIANTS
4858 	if (ndp->ni_pathlen != fpl->debug.ni_pathlen) {
4859 		panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n",
4860 		    __func__, ndp->ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar,
4861 		    cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf);
4862 	}
4863 #endif
4864 
4865 	if (__predict_false(len + ndp->ni_pathlen > MAXPATHLEN)) {
4866 		return (ENAMETOOLONG);
4867 	}
4868 
4869 	if (__predict_false(ndp->ni_loopcnt++ >= MAXSYMLINKS)) {
4870 		return (ELOOP);
4871 	}
4872 
4873 	adjust = len;
4874 	if (ndp->ni_pathlen > 1) {
4875 		bcopy(ndp->ni_next, cnp->cn_pnbuf + len, ndp->ni_pathlen);
4876 	} else {
4877 		if (cache_fpl_istrailingslash(fpl)) {
4878 			adjust = len + 1;
4879 			cnp->cn_pnbuf[len] = '/';
4880 			cnp->cn_pnbuf[len + 1] = '\0';
4881 		} else {
4882 			cnp->cn_pnbuf[len] = '\0';
4883 		}
4884 	}
4885 	bcopy(string, cnp->cn_pnbuf, len);
4886 
4887 	ndp->ni_pathlen += adjust;
4888 	cache_fpl_pathlen_add(fpl, adjust);
4889 	cnp->cn_nameptr = cnp->cn_pnbuf;
4890 	fpl->nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1];
4891 	fpl->tvp = NULL;
4892 	return (0);
4893 }
4894 
4895 static int __noinline
4896 cache_fplookup_symlink(struct cache_fpl *fpl)
4897 {
4898 	struct mount *mp;
4899 	struct nameidata *ndp;
4900 	struct componentname *cnp;
4901 	struct vnode *dvp, *tvp;
4902 	int error;
4903 
4904 	ndp = fpl->ndp;
4905 	cnp = fpl->cnp;
4906 	dvp = fpl->dvp;
4907 	tvp = fpl->tvp;
4908 
4909 	if (cache_fpl_islastcn(ndp)) {
4910 		if ((cnp->cn_flags & FOLLOW) == 0) {
4911 			return (cache_fplookup_final(fpl));
4912 		}
4913 	}
4914 
4915 	mp = atomic_load_ptr(&dvp->v_mount);
4916 	if (__predict_false(mp == NULL)) {
4917 		return (cache_fpl_aborted(fpl));
4918 	}
4919 
4920 	/*
4921 	 * Note this check races against setting the flag just like regular
4922 	 * lookup.
4923 	 */
4924 	if (__predict_false((mp->mnt_flag & MNT_NOSYMFOLLOW) != 0)) {
4925 		cache_fpl_smr_exit(fpl);
4926 		return (cache_fpl_handled_error(fpl, EACCES));
4927 	}
4928 
4929 	error = VOP_FPLOOKUP_SYMLINK(tvp, fpl);
4930 	if (__predict_false(error != 0)) {
4931 		switch (error) {
4932 		case EAGAIN:
4933 			return (cache_fpl_partial(fpl));
4934 		case ENOENT:
4935 		case ENAMETOOLONG:
4936 		case ELOOP:
4937 			cache_fpl_smr_exit(fpl);
4938 			return (cache_fpl_handled_error(fpl, error));
4939 		default:
4940 			return (cache_fpl_aborted(fpl));
4941 		}
4942 	}
4943 
4944 	if (*(cnp->cn_nameptr) == '/') {
4945 		fpl->dvp = cache_fpl_handle_root(fpl);
4946 		fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp);
4947 		if (seqc_in_modify(fpl->dvp_seqc)) {
4948 			return (cache_fpl_aborted(fpl));
4949 		}
4950 	}
4951 	return (0);
4952 }
4953 
4954 static int
4955 cache_fplookup_next(struct cache_fpl *fpl)
4956 {
4957 	struct componentname *cnp;
4958 	struct namecache *ncp;
4959 	struct vnode *dvp, *tvp;
4960 	u_char nc_flag;
4961 	uint32_t hash;
4962 	int error;
4963 
4964 	cnp = fpl->cnp;
4965 	dvp = fpl->dvp;
4966 	hash = fpl->hash;
4967 
4968 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
4969 		if (cnp->cn_namelen == 1) {
4970 			return (cache_fplookup_dot(fpl));
4971 		}
4972 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
4973 			return (cache_fplookup_dotdot(fpl));
4974 		}
4975 	}
4976 
4977 	MPASS(!cache_fpl_isdotdot(cnp));
4978 
4979 	CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
4980 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
4981 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
4982 			break;
4983 	}
4984 
4985 	if (__predict_false(ncp == NULL)) {
4986 		return (cache_fplookup_noentry(fpl));
4987 	}
4988 
4989 	tvp = atomic_load_ptr(&ncp->nc_vp);
4990 	nc_flag = atomic_load_char(&ncp->nc_flag);
4991 	if ((nc_flag & NCF_NEGATIVE) != 0) {
4992 		return (cache_fplookup_neg(fpl, ncp, hash));
4993 	}
4994 
4995 	if (!cache_ncp_canuse(ncp)) {
4996 		return (cache_fpl_partial(fpl));
4997 	}
4998 
4999 	fpl->tvp = tvp;
5000 	fpl->tvp_seqc = vn_seqc_read_any(tvp);
5001 	if (seqc_in_modify(fpl->tvp_seqc)) {
5002 		return (cache_fpl_partial(fpl));
5003 	}
5004 
5005 	counter_u64_add(numposhits, 1);
5006 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, tvp);
5007 
5008 	error = 0;
5009 	if (cache_fplookup_is_mp(fpl)) {
5010 		error = cache_fplookup_cross_mount(fpl);
5011 	}
5012 	return (error);
5013 }
5014 
5015 static bool
5016 cache_fplookup_mp_supported(struct mount *mp)
5017 {
5018 
5019 	MPASS(mp != NULL);
5020 	if ((mp->mnt_kern_flag & MNTK_FPLOOKUP) == 0)
5021 		return (false);
5022 	return (true);
5023 }
5024 
5025 /*
5026  * Walk up the mount stack (if any).
5027  *
5028  * Correctness is provided in the following ways:
5029  * - all vnodes are protected from freeing with SMR
5030  * - struct mount objects are type stable making them always safe to access
5031  * - stability of the particular mount is provided by busying it
5032  * - relationship between the vnode which is mounted on and the mount is
5033  *   verified with the vnode sequence counter after busying
5034  * - association between root vnode of the mount and the mount is protected
5035  *   by busy
5036  *
5037  * From that point on we can read the sequence counter of the root vnode
5038  * and get the next mount on the stack (if any) using the same protection.
5039  *
5040  * By the end of successful walk we are guaranteed the reached state was
5041  * indeed present at least at some point which matches the regular lookup.
5042  */
5043 static int __noinline
5044 cache_fplookup_climb_mount(struct cache_fpl *fpl)
5045 {
5046 	struct mount *mp, *prev_mp;
5047 	struct mount_pcpu *mpcpu, *prev_mpcpu;
5048 	struct vnode *vp;
5049 	seqc_t vp_seqc;
5050 
5051 	vp = fpl->tvp;
5052 	vp_seqc = fpl->tvp_seqc;
5053 
5054 	VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp);
5055 	mp = atomic_load_ptr(&vp->v_mountedhere);
5056 	if (__predict_false(mp == NULL)) {
5057 		return (0);
5058 	}
5059 
5060 	prev_mp = NULL;
5061 	for (;;) {
5062 		if (!vfs_op_thread_enter_crit(mp, mpcpu)) {
5063 			if (prev_mp != NULL)
5064 				vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
5065 			return (cache_fpl_partial(fpl));
5066 		}
5067 		if (prev_mp != NULL)
5068 			vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
5069 		if (!vn_seqc_consistent(vp, vp_seqc)) {
5070 			vfs_op_thread_exit_crit(mp, mpcpu);
5071 			return (cache_fpl_partial(fpl));
5072 		}
5073 		if (!cache_fplookup_mp_supported(mp)) {
5074 			vfs_op_thread_exit_crit(mp, mpcpu);
5075 			return (cache_fpl_partial(fpl));
5076 		}
5077 		vp = atomic_load_ptr(&mp->mnt_rootvnode);
5078 		if (vp == NULL) {
5079 			vfs_op_thread_exit_crit(mp, mpcpu);
5080 			return (cache_fpl_partial(fpl));
5081 		}
5082 		vp_seqc = vn_seqc_read_any(vp);
5083 		if (seqc_in_modify(vp_seqc)) {
5084 			vfs_op_thread_exit_crit(mp, mpcpu);
5085 			return (cache_fpl_partial(fpl));
5086 		}
5087 		prev_mp = mp;
5088 		prev_mpcpu = mpcpu;
5089 		mp = atomic_load_ptr(&vp->v_mountedhere);
5090 		if (mp == NULL)
5091 			break;
5092 	}
5093 
5094 	vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
5095 	fpl->tvp = vp;
5096 	fpl->tvp_seqc = vp_seqc;
5097 	return (0);
5098 }
5099 
5100 static int __noinline
5101 cache_fplookup_cross_mount(struct cache_fpl *fpl)
5102 {
5103 	struct mount *mp;
5104 	struct mount_pcpu *mpcpu;
5105 	struct vnode *vp;
5106 	seqc_t vp_seqc;
5107 
5108 	vp = fpl->tvp;
5109 	vp_seqc = fpl->tvp_seqc;
5110 
5111 	VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp);
5112 	mp = atomic_load_ptr(&vp->v_mountedhere);
5113 	if (__predict_false(mp == NULL)) {
5114 		return (0);
5115 	}
5116 
5117 	if (!vfs_op_thread_enter_crit(mp, mpcpu)) {
5118 		return (cache_fpl_partial(fpl));
5119 	}
5120 	if (!vn_seqc_consistent(vp, vp_seqc)) {
5121 		vfs_op_thread_exit_crit(mp, mpcpu);
5122 		return (cache_fpl_partial(fpl));
5123 	}
5124 	if (!cache_fplookup_mp_supported(mp)) {
5125 		vfs_op_thread_exit_crit(mp, mpcpu);
5126 		return (cache_fpl_partial(fpl));
5127 	}
5128 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
5129 	if (__predict_false(vp == NULL)) {
5130 		vfs_op_thread_exit_crit(mp, mpcpu);
5131 		return (cache_fpl_partial(fpl));
5132 	}
5133 	vp_seqc = vn_seqc_read_any(vp);
5134 	vfs_op_thread_exit_crit(mp, mpcpu);
5135 	if (seqc_in_modify(vp_seqc)) {
5136 		return (cache_fpl_partial(fpl));
5137 	}
5138 	mp = atomic_load_ptr(&vp->v_mountedhere);
5139 	if (__predict_false(mp != NULL)) {
5140 		/*
5141 		 * There are possibly more mount points on top.
5142 		 * Normally this does not happen so for simplicity just start
5143 		 * over.
5144 		 */
5145 		return (cache_fplookup_climb_mount(fpl));
5146 	}
5147 
5148 	fpl->tvp = vp;
5149 	fpl->tvp_seqc = vp_seqc;
5150 	return (0);
5151 }
5152 
5153 /*
5154  * Check if a vnode is mounted on.
5155  */
5156 static bool
5157 cache_fplookup_is_mp(struct cache_fpl *fpl)
5158 {
5159 	struct vnode *vp;
5160 
5161 	vp = fpl->tvp;
5162 	return ((vn_irflag_read(vp) & VIRF_MOUNTPOINT) != 0);
5163 }
5164 
5165 /*
5166  * Parse the path.
5167  *
5168  * The code was originally copy-pasted from regular lookup and despite
5169  * clean ups leaves performance on the table. Any modifications here
5170  * must take into account that in case off fallback the resulting
5171  * nameidata state has to be compatible with the original.
5172  */
5173 
5174 /*
5175  * Debug ni_pathlen tracking.
5176  */
5177 #ifdef INVARIANTS
5178 static void
5179 cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n)
5180 {
5181 
5182 	fpl->debug.ni_pathlen += n;
5183 	KASSERT(fpl->debug.ni_pathlen <= PATH_MAX,
5184 	    ("%s: pathlen overflow to %zd\n", __func__, fpl->debug.ni_pathlen));
5185 }
5186 
5187 static void
5188 cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n)
5189 {
5190 
5191 	fpl->debug.ni_pathlen -= n;
5192 	KASSERT(fpl->debug.ni_pathlen <= PATH_MAX,
5193 	    ("%s: pathlen underflow to %zd\n", __func__, fpl->debug.ni_pathlen));
5194 }
5195 
5196 static void
5197 cache_fpl_pathlen_inc(struct cache_fpl *fpl)
5198 {
5199 
5200 	cache_fpl_pathlen_add(fpl, 1);
5201 }
5202 
5203 static void
5204 cache_fpl_pathlen_dec(struct cache_fpl *fpl)
5205 {
5206 
5207 	cache_fpl_pathlen_sub(fpl, 1);
5208 }
5209 #else
5210 static void
5211 cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n)
5212 {
5213 }
5214 
5215 static void
5216 cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n)
5217 {
5218 }
5219 
5220 static void
5221 cache_fpl_pathlen_inc(struct cache_fpl *fpl)
5222 {
5223 }
5224 
5225 static void
5226 cache_fpl_pathlen_dec(struct cache_fpl *fpl)
5227 {
5228 }
5229 #endif
5230 
5231 static void
5232 cache_fplookup_parse(struct cache_fpl *fpl)
5233 {
5234 	struct nameidata *ndp;
5235 	struct componentname *cnp;
5236 	struct vnode *dvp;
5237 	char *cp;
5238 	uint32_t hash;
5239 
5240 	ndp = fpl->ndp;
5241 	cnp = fpl->cnp;
5242 	dvp = fpl->dvp;
5243 
5244 	/*
5245 	 * Find the end of this path component, it is either / or nul.
5246 	 *
5247 	 * Store / as a temporary sentinel so that we only have one character
5248 	 * to test for. Pathnames tend to be short so this should not be
5249 	 * resulting in cache misses.
5250 	 *
5251 	 * TODO: fix this to be word-sized.
5252 	 */
5253 	KASSERT(&cnp->cn_nameptr[fpl->debug.ni_pathlen - 1] == fpl->nulchar,
5254 	    ("%s: mismatch between pathlen (%zu) and nulchar (%p != %p), string [%s]\n",
5255 	    __func__, fpl->debug.ni_pathlen, &cnp->cn_nameptr[fpl->debug.ni_pathlen - 1],
5256 	    fpl->nulchar, cnp->cn_pnbuf));
5257 	KASSERT(*fpl->nulchar == '\0',
5258 	    ("%s: expected nul at %p; string [%s]\n", __func__, fpl->nulchar,
5259 	    cnp->cn_pnbuf));
5260 	hash = cache_get_hash_iter_start(dvp);
5261 	*fpl->nulchar = '/';
5262 	for (cp = cnp->cn_nameptr; *cp != '/'; cp++) {
5263 		KASSERT(*cp != '\0',
5264 		    ("%s: encountered unexpected nul; string [%s]\n", __func__,
5265 		    cnp->cn_nameptr));
5266 		hash = cache_get_hash_iter(*cp, hash);
5267 		continue;
5268 	}
5269 	*fpl->nulchar = '\0';
5270 	fpl->hash = cache_get_hash_iter_finish(hash);
5271 
5272 	cnp->cn_namelen = cp - cnp->cn_nameptr;
5273 	cache_fpl_pathlen_sub(fpl, cnp->cn_namelen);
5274 
5275 #ifdef INVARIANTS
5276 	/*
5277 	 * cache_get_hash only accepts lengths up to NAME_MAX. This is fine since
5278 	 * we are going to fail this lookup with ENAMETOOLONG (see below).
5279 	 */
5280 	if (cnp->cn_namelen <= NAME_MAX) {
5281 		if (fpl->hash != cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp)) {
5282 			panic("%s: mismatched hash for [%s] len %ld", __func__,
5283 			    cnp->cn_nameptr, cnp->cn_namelen);
5284 		}
5285 	}
5286 #endif
5287 
5288 	/*
5289 	 * Hack: we have to check if the found path component's length exceeds
5290 	 * NAME_MAX. However, the condition is very rarely true and check can
5291 	 * be elided in the common case -- if an entry was found in the cache,
5292 	 * then it could not have been too long to begin with.
5293 	 */
5294 	ndp->ni_next = cp;
5295 }
5296 
5297 static void
5298 cache_fplookup_parse_advance(struct cache_fpl *fpl)
5299 {
5300 	struct nameidata *ndp;
5301 	struct componentname *cnp;
5302 
5303 	ndp = fpl->ndp;
5304 	cnp = fpl->cnp;
5305 
5306 	cnp->cn_nameptr = ndp->ni_next;
5307 	KASSERT(*(cnp->cn_nameptr) == '/',
5308 	    ("%s: should have seen slash at %p ; buf %p [%s]\n", __func__,
5309 	    cnp->cn_nameptr, cnp->cn_pnbuf, cnp->cn_pnbuf));
5310 	cnp->cn_nameptr++;
5311 	cache_fpl_pathlen_dec(fpl);
5312 }
5313 
5314 /*
5315  * Skip spurious slashes in a pathname (e.g., "foo///bar") and retry.
5316  *
5317  * Lockless lookup tries to elide checking for spurious slashes and should they
5318  * be present is guaranteed to fail to find an entry. In this case the caller
5319  * must check if the name starts with a slash and call this routine.  It is
5320  * going to fast forward across the spurious slashes and set the state up for
5321  * retry.
5322  */
5323 static int __noinline
5324 cache_fplookup_skip_slashes(struct cache_fpl *fpl)
5325 {
5326 	struct nameidata *ndp;
5327 	struct componentname *cnp;
5328 
5329 	ndp = fpl->ndp;
5330 	cnp = fpl->cnp;
5331 
5332 	MPASS(*(cnp->cn_nameptr) == '/');
5333 	do {
5334 		cnp->cn_nameptr++;
5335 		cache_fpl_pathlen_dec(fpl);
5336 	} while (*(cnp->cn_nameptr) == '/');
5337 
5338 	/*
5339 	 * Go back to one slash so that cache_fplookup_parse_advance has
5340 	 * something to skip.
5341 	 */
5342 	cnp->cn_nameptr--;
5343 	cache_fpl_pathlen_inc(fpl);
5344 
5345 	/*
5346 	 * cache_fplookup_parse_advance starts from ndp->ni_next
5347 	 */
5348 	ndp->ni_next = cnp->cn_nameptr;
5349 
5350 	/*
5351 	 * See cache_fplookup_dot.
5352 	 */
5353 	fpl->tvp = fpl->dvp;
5354 	fpl->tvp_seqc = fpl->dvp_seqc;
5355 
5356 	return (0);
5357 }
5358 
5359 /*
5360  * Handle trailing slashes (e.g., "foo/").
5361  *
5362  * If a trailing slash is found the terminal vnode must be a directory.
5363  * Regular lookup shortens the path by nulifying the first trailing slash and
5364  * sets the TRAILINGSLASH flag to denote this took place. There are several
5365  * checks on it performed later.
5366  *
5367  * Similarly to spurious slashes, lockless lookup handles this in a speculative
5368  * manner relying on an invariant that a non-directory vnode will get a miss.
5369  * In this case cn_nameptr[0] == '\0' and cn_namelen == 0.
5370  *
5371  * Thus for a path like "foo/bar/" the code unwinds the state back to "bar/"
5372  * and denotes this is the last path component, which avoids looping back.
5373  *
5374  * Only plain lookups are supported for now to restrict corner cases to handle.
5375  */
5376 static int __noinline
5377 cache_fplookup_trailingslash(struct cache_fpl *fpl)
5378 {
5379 #ifdef INVARIANTS
5380 	size_t ni_pathlen;
5381 #endif
5382 	struct nameidata *ndp;
5383 	struct componentname *cnp;
5384 	struct namecache *ncp;
5385 	struct vnode *tvp;
5386 	char *cn_nameptr_orig, *cn_nameptr_slash;
5387 	seqc_t tvp_seqc;
5388 	u_char nc_flag;
5389 
5390 	ndp = fpl->ndp;
5391 	cnp = fpl->cnp;
5392 	tvp = fpl->tvp;
5393 	tvp_seqc = fpl->tvp_seqc;
5394 
5395 	MPASS(fpl->dvp == fpl->tvp);
5396 	KASSERT(cache_fpl_istrailingslash(fpl),
5397 	    ("%s: expected trailing slash at %p; string [%s]\n", __func__, fpl->nulchar - 1,
5398 	    cnp->cn_pnbuf));
5399 	KASSERT(cnp->cn_nameptr[0] == '\0',
5400 	    ("%s: expected nul char at %p; string [%s]\n", __func__, &cnp->cn_nameptr[0],
5401 	    cnp->cn_pnbuf));
5402 	KASSERT(cnp->cn_namelen == 0,
5403 	    ("%s: namelen 0 but got %ld; string [%s]\n", __func__, cnp->cn_namelen,
5404 	    cnp->cn_pnbuf));
5405 	MPASS(cnp->cn_nameptr > cnp->cn_pnbuf);
5406 
5407 	if (cnp->cn_nameiop != LOOKUP) {
5408 		return (cache_fpl_aborted(fpl));
5409 	}
5410 
5411 	if (__predict_false(tvp->v_type != VDIR)) {
5412 		if (!vn_seqc_consistent(tvp, tvp_seqc)) {
5413 			return (cache_fpl_aborted(fpl));
5414 		}
5415 		cache_fpl_smr_exit(fpl);
5416 		return (cache_fpl_handled_error(fpl, ENOTDIR));
5417 	}
5418 
5419 	/*
5420 	 * Denote the last component.
5421 	 */
5422 	ndp->ni_next = &cnp->cn_nameptr[0];
5423 	MPASS(cache_fpl_islastcn(ndp));
5424 
5425 	/*
5426 	 * Unwind trailing slashes.
5427 	 */
5428 	cn_nameptr_orig = cnp->cn_nameptr;
5429 	while (cnp->cn_nameptr >= cnp->cn_pnbuf) {
5430 		cnp->cn_nameptr--;
5431 		if (cnp->cn_nameptr[0] != '/') {
5432 			break;
5433 		}
5434 	}
5435 
5436 	/*
5437 	 * Unwind to the beginning of the path component.
5438 	 *
5439 	 * Note the path may or may not have started with a slash.
5440 	 */
5441 	cn_nameptr_slash = cnp->cn_nameptr;
5442 	while (cnp->cn_nameptr > cnp->cn_pnbuf) {
5443 		cnp->cn_nameptr--;
5444 		if (cnp->cn_nameptr[0] == '/') {
5445 			break;
5446 		}
5447 	}
5448 	if (cnp->cn_nameptr[0] == '/') {
5449 		cnp->cn_nameptr++;
5450 	}
5451 
5452 	cnp->cn_namelen = cn_nameptr_slash - cnp->cn_nameptr + 1;
5453 	cache_fpl_pathlen_add(fpl, cn_nameptr_orig - cnp->cn_nameptr);
5454 	cache_fpl_checkpoint(fpl);
5455 
5456 #ifdef INVARIANTS
5457 	ni_pathlen = fpl->nulchar - cnp->cn_nameptr + 1;
5458 	if (ni_pathlen != fpl->debug.ni_pathlen) {
5459 		panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n",
5460 		    __func__, ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar,
5461 		    cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf);
5462 	}
5463 #endif
5464 
5465 	/*
5466 	 * If this was a "./" lookup the parent directory is already correct.
5467 	 */
5468 	if (cnp->cn_nameptr[0] == '.' && cnp->cn_namelen == 1) {
5469 		return (0);
5470 	}
5471 
5472 	/*
5473 	 * Otherwise we need to look it up.
5474 	 */
5475 	tvp = fpl->tvp;
5476 	ncp = atomic_load_consume_ptr(&tvp->v_cache_dd);
5477 	if (__predict_false(ncp == NULL)) {
5478 		return (cache_fpl_aborted(fpl));
5479 	}
5480 	nc_flag = atomic_load_char(&ncp->nc_flag);
5481 	if ((nc_flag & NCF_ISDOTDOT) != 0) {
5482 		return (cache_fpl_aborted(fpl));
5483 	}
5484 	fpl->dvp = ncp->nc_dvp;
5485 	fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp);
5486 	if (seqc_in_modify(fpl->dvp_seqc)) {
5487 		return (cache_fpl_aborted(fpl));
5488 	}
5489 	return (0);
5490 }
5491 
5492 /*
5493  * See the API contract for VOP_FPLOOKUP_VEXEC.
5494  */
5495 static int __noinline
5496 cache_fplookup_failed_vexec(struct cache_fpl *fpl, int error)
5497 {
5498 	struct componentname *cnp;
5499 	struct vnode *dvp;
5500 	seqc_t dvp_seqc;
5501 
5502 	cnp = fpl->cnp;
5503 	dvp = fpl->dvp;
5504 	dvp_seqc = fpl->dvp_seqc;
5505 
5506 	/*
5507 	 * TODO: Due to ignoring trailing slashes lookup will perform a
5508 	 * permission check on the last dir when it should not be doing it.  It
5509 	 * may fail, but said failure should be ignored. It is possible to fix
5510 	 * it up fully without resorting to regular lookup, but for now just
5511 	 * abort.
5512 	 */
5513 	if (cache_fpl_istrailingslash(fpl)) {
5514 		return (cache_fpl_aborted(fpl));
5515 	}
5516 
5517 	/*
5518 	 * Hack: delayed degenerate path checking.
5519 	 */
5520 	if (cnp->cn_nameptr[0] == '\0' && fpl->tvp == NULL) {
5521 		return (cache_fplookup_degenerate(fpl));
5522 	}
5523 
5524 	/*
5525 	 * Hack: delayed name len checking.
5526 	 */
5527 	if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
5528 		cache_fpl_smr_exit(fpl);
5529 		return (cache_fpl_handled_error(fpl, ENAMETOOLONG));
5530 	}
5531 
5532 	/*
5533 	 * Hack: they may be looking up foo/bar, where foo is not a directory.
5534 	 * In such a case we need to return ENOTDIR, but we may happen to get
5535 	 * here with a different error.
5536 	 */
5537 	if (dvp->v_type != VDIR) {
5538 		error = ENOTDIR;
5539 	}
5540 
5541 	/*
5542 	 * Hack: handle O_SEARCH.
5543 	 *
5544 	 * Open Group Base Specifications Issue 7, 2018 edition states:
5545 	 * <quote>
5546 	 * If the access mode of the open file description associated with the
5547 	 * file descriptor is not O_SEARCH, the function shall check whether
5548 	 * directory searches are permitted using the current permissions of
5549 	 * the directory underlying the file descriptor. If the access mode is
5550 	 * O_SEARCH, the function shall not perform the check.
5551 	 * </quote>
5552 	 *
5553 	 * Regular lookup tests for the NOEXECCHECK flag for every path
5554 	 * component to decide whether to do the permission check. However,
5555 	 * since most lookups never have the flag (and when they do it is only
5556 	 * present for the first path component), lockless lookup only acts on
5557 	 * it if there is a permission problem. Here the flag is represented
5558 	 * with a boolean so that we don't have to clear it on the way out.
5559 	 *
5560 	 * For simplicity this always aborts.
5561 	 * TODO: check if this is the first lookup and ignore the permission
5562 	 * problem. Note the flag has to survive fallback (if it happens to be
5563 	 * performed).
5564 	 */
5565 	if (fpl->fsearch) {
5566 		return (cache_fpl_aborted(fpl));
5567 	}
5568 
5569 	switch (error) {
5570 	case EAGAIN:
5571 		if (!vn_seqc_consistent(dvp, dvp_seqc)) {
5572 			error = cache_fpl_aborted(fpl);
5573 		} else {
5574 			cache_fpl_partial(fpl);
5575 		}
5576 		break;
5577 	default:
5578 		if (!vn_seqc_consistent(dvp, dvp_seqc)) {
5579 			error = cache_fpl_aborted(fpl);
5580 		} else {
5581 			cache_fpl_smr_exit(fpl);
5582 			cache_fpl_handled_error(fpl, error);
5583 		}
5584 		break;
5585 	}
5586 	return (error);
5587 }
5588 
5589 static int
5590 cache_fplookup_impl(struct vnode *dvp, struct cache_fpl *fpl)
5591 {
5592 	struct nameidata *ndp;
5593 	struct componentname *cnp;
5594 	struct mount *mp;
5595 	int error;
5596 
5597 	ndp = fpl->ndp;
5598 	cnp = fpl->cnp;
5599 
5600 	cache_fpl_checkpoint(fpl);
5601 
5602 	/*
5603 	 * The vnode at hand is almost always stable, skip checking for it.
5604 	 * Worst case this postpones the check towards the end of the iteration
5605 	 * of the main loop.
5606 	 */
5607 	fpl->dvp = dvp;
5608 	fpl->dvp_seqc = vn_seqc_read_notmodify(fpl->dvp);
5609 
5610 	mp = atomic_load_ptr(&dvp->v_mount);
5611 	if (__predict_false(mp == NULL || !cache_fplookup_mp_supported(mp))) {
5612 		return (cache_fpl_aborted(fpl));
5613 	}
5614 
5615 	MPASS(fpl->tvp == NULL);
5616 
5617 	for (;;) {
5618 		cache_fplookup_parse(fpl);
5619 
5620 		error = VOP_FPLOOKUP_VEXEC(fpl->dvp, cnp->cn_cred);
5621 		if (__predict_false(error != 0)) {
5622 			error = cache_fplookup_failed_vexec(fpl, error);
5623 			break;
5624 		}
5625 
5626 		error = cache_fplookup_next(fpl);
5627 		if (__predict_false(cache_fpl_terminated(fpl))) {
5628 			break;
5629 		}
5630 
5631 		VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp);
5632 
5633 		if (fpl->tvp->v_type == VLNK) {
5634 			error = cache_fplookup_symlink(fpl);
5635 			if (cache_fpl_terminated(fpl)) {
5636 				break;
5637 			}
5638 		} else {
5639 			if (cache_fpl_islastcn(ndp)) {
5640 				error = cache_fplookup_final(fpl);
5641 				break;
5642 			}
5643 
5644 			if (!vn_seqc_consistent(fpl->dvp, fpl->dvp_seqc)) {
5645 				error = cache_fpl_aborted(fpl);
5646 				break;
5647 			}
5648 
5649 			fpl->dvp = fpl->tvp;
5650 			fpl->dvp_seqc = fpl->tvp_seqc;
5651 			cache_fplookup_parse_advance(fpl);
5652 		}
5653 
5654 		cache_fpl_checkpoint(fpl);
5655 	}
5656 
5657 	return (error);
5658 }
5659 
5660 /*
5661  * Fast path lookup protected with SMR and sequence counters.
5662  *
5663  * Note: all VOP_FPLOOKUP_VEXEC routines have a comment referencing this one.
5664  *
5665  * Filesystems can opt in by setting the MNTK_FPLOOKUP flag and meeting criteria
5666  * outlined below.
5667  *
5668  * Traditional vnode lookup conceptually looks like this:
5669  *
5670  * vn_lock(current);
5671  * for (;;) {
5672  *	next = find();
5673  *	vn_lock(next);
5674  *	vn_unlock(current);
5675  *	current = next;
5676  *	if (last)
5677  *	    break;
5678  * }
5679  * return (current);
5680  *
5681  * Each jump to the next vnode is safe memory-wise and atomic with respect to
5682  * any modifications thanks to holding respective locks.
5683  *
5684  * The same guarantee can be provided with a combination of safe memory
5685  * reclamation and sequence counters instead. If all operations which affect
5686  * the relationship between the current vnode and the one we are looking for
5687  * also modify the counter, we can verify whether all the conditions held as
5688  * we made the jump. This includes things like permissions, mount points etc.
5689  * Counter modification is provided by enclosing relevant places in
5690  * vn_seqc_write_begin()/end() calls.
5691  *
5692  * Thus this translates to:
5693  *
5694  * vfs_smr_enter();
5695  * dvp_seqc = seqc_read_any(dvp);
5696  * if (seqc_in_modify(dvp_seqc)) // someone is altering the vnode
5697  *     abort();
5698  * for (;;) {
5699  * 	tvp = find();
5700  * 	tvp_seqc = seqc_read_any(tvp);
5701  * 	if (seqc_in_modify(tvp_seqc)) // someone is altering the target vnode
5702  * 	    abort();
5703  * 	if (!seqc_consistent(dvp, dvp_seqc) // someone is altering the vnode
5704  * 	    abort();
5705  * 	dvp = tvp; // we know nothing of importance has changed
5706  * 	dvp_seqc = tvp_seqc; // store the counter for the tvp iteration
5707  * 	if (last)
5708  * 	    break;
5709  * }
5710  * vget(); // secure the vnode
5711  * if (!seqc_consistent(tvp, tvp_seqc) // final check
5712  * 	    abort();
5713  * // at this point we know nothing has changed for any parent<->child pair
5714  * // as they were crossed during the lookup, meaning we matched the guarantee
5715  * // of the locked variant
5716  * return (tvp);
5717  *
5718  * The API contract for VOP_FPLOOKUP_VEXEC routines is as follows:
5719  * - they are called while within vfs_smr protection which they must never exit
5720  * - EAGAIN can be returned to denote checking could not be performed, it is
5721  *   always valid to return it
5722  * - if the sequence counter has not changed the result must be valid
5723  * - if the sequence counter has changed both false positives and false negatives
5724  *   are permitted (since the result will be rejected later)
5725  * - for simple cases of unix permission checks vaccess_vexec_smr can be used
5726  *
5727  * Caveats to watch out for:
5728  * - vnodes are passed unlocked and unreferenced with nothing stopping
5729  *   VOP_RECLAIM, in turn meaning that ->v_data can become NULL. It is advised
5730  *   to use atomic_load_ptr to fetch it.
5731  * - the aforementioned object can also get freed, meaning absent other means it
5732  *   should be protected with vfs_smr
5733  * - either safely checking permissions as they are modified or guaranteeing
5734  *   their stability is left to the routine
5735  */
5736 int
5737 cache_fplookup(struct nameidata *ndp, enum cache_fpl_status *status,
5738     struct pwd **pwdp)
5739 {
5740 	struct cache_fpl fpl;
5741 	struct pwd *pwd;
5742 	struct vnode *dvp;
5743 	struct componentname *cnp;
5744 	int error;
5745 
5746 	fpl.status = CACHE_FPL_STATUS_UNSET;
5747 	fpl.in_smr = false;
5748 	fpl.ndp = ndp;
5749 	fpl.cnp = cnp = &ndp->ni_cnd;
5750 	MPASS(ndp->ni_lcf == 0);
5751 	MPASS(curthread == cnp->cn_thread);
5752 	KASSERT ((cnp->cn_flags & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
5753 	    ("%s: internal flags found in cn_flags %" PRIx64, __func__,
5754 	    cnp->cn_flags));
5755 	if ((cnp->cn_flags & SAVESTART) != 0) {
5756 		MPASS(cnp->cn_nameiop != LOOKUP);
5757 	}
5758 	MPASS(cnp->cn_nameptr == cnp->cn_pnbuf);
5759 
5760 	if (__predict_false(!cache_can_fplookup(&fpl))) {
5761 		*status = fpl.status;
5762 		SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
5763 		return (EOPNOTSUPP);
5764 	}
5765 
5766 	cache_fpl_checkpoint_outer(&fpl);
5767 
5768 	cache_fpl_smr_enter_initial(&fpl);
5769 #ifdef INVARIANTS
5770 	fpl.debug.ni_pathlen = ndp->ni_pathlen;
5771 #endif
5772 	fpl.nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1];
5773 	fpl.fsearch = false;
5774 	fpl.savename = (cnp->cn_flags & SAVENAME) != 0;
5775 	fpl.tvp = NULL; /* for degenerate path handling */
5776 	fpl.pwd = pwdp;
5777 	pwd = pwd_get_smr();
5778 	*(fpl.pwd) = pwd;
5779 	ndp->ni_rootdir = pwd->pwd_rdir;
5780 	ndp->ni_topdir = pwd->pwd_jdir;
5781 
5782 	if (cnp->cn_pnbuf[0] == '/') {
5783 		dvp = cache_fpl_handle_root(&fpl);
5784 		MPASS(ndp->ni_resflags == 0);
5785 		ndp->ni_resflags = NIRES_ABS;
5786 	} else {
5787 		if (ndp->ni_dirfd == AT_FDCWD) {
5788 			dvp = pwd->pwd_cdir;
5789 		} else {
5790 			error = cache_fplookup_dirfd(&fpl, &dvp);
5791 			if (__predict_false(error != 0)) {
5792 				goto out;
5793 			}
5794 		}
5795 	}
5796 
5797 	SDT_PROBE4(vfs, namei, lookup, entry, dvp, cnp->cn_pnbuf, cnp->cn_flags, true);
5798 	error = cache_fplookup_impl(dvp, &fpl);
5799 out:
5800 	cache_fpl_smr_assert_not_entered(&fpl);
5801 	cache_fpl_assert_status(&fpl);
5802 	*status = fpl.status;
5803 	if (SDT_PROBES_ENABLED()) {
5804 		SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
5805 		if (fpl.status == CACHE_FPL_STATUS_HANDLED)
5806 			SDT_PROBE4(vfs, namei, lookup, return, error, ndp->ni_vp, true,
5807 			    ndp);
5808 	}
5809 
5810 	if (__predict_true(fpl.status == CACHE_FPL_STATUS_HANDLED)) {
5811 		MPASS(error != CACHE_FPL_FAILED);
5812 		if (error != 0) {
5813 			MPASS(fpl.dvp == NULL);
5814 			MPASS(fpl.tvp == NULL);
5815 			MPASS(fpl.savename == false);
5816 		}
5817 		ndp->ni_dvp = fpl.dvp;
5818 		ndp->ni_vp = fpl.tvp;
5819 		if (fpl.savename) {
5820 			cnp->cn_flags |= HASBUF;
5821 		} else {
5822 			cache_fpl_cleanup_cnp(cnp);
5823 		}
5824 	}
5825 	return (error);
5826 }
5827