xref: /freebsd/sys/kern/vfs_cache.c (revision 82397d791966b09d344251bc709cd9db2b3a1902)
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 
126 SDT_PROBE_DEFINE3(vfs, fplookup, lookup, done, "struct nameidata", "int", "bool");
127 SDT_PROBE_DECLARE(vfs, namei, lookup, entry);
128 SDT_PROBE_DECLARE(vfs, namei, lookup, return);
129 
130 /*
131  * This structure describes the elements in the cache of recent
132  * names looked up by namei.
133  */
134 struct negstate {
135 	u_char neg_flag;
136 	u_char neg_hit;
137 };
138 _Static_assert(sizeof(struct negstate) <= sizeof(struct vnode *),
139     "the state must fit in a union with a pointer without growing it");
140 
141 struct	namecache {
142 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
143 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
144 	CK_SLIST_ENTRY(namecache) nc_hash;/* hash chain */
145 	struct	vnode *nc_dvp;		/* vnode of parent of name */
146 	union {
147 		struct	vnode *nu_vp;	/* vnode the name refers to */
148 		struct	negstate nu_neg;/* negative entry state */
149 	} n_un;
150 	u_char	nc_flag;		/* flag bits */
151 	u_char	nc_nlen;		/* length of name */
152 	char	nc_name[0];		/* segment name + nul */
153 };
154 
155 /*
156  * struct namecache_ts repeats struct namecache layout up to the
157  * nc_nlen member.
158  * struct namecache_ts is used in place of struct namecache when time(s) need
159  * to be stored.  The nc_dotdottime field is used when a cache entry is mapping
160  * both a non-dotdot directory name plus dotdot for the directory's
161  * parent.
162  *
163  * See below for alignment requirement.
164  */
165 struct	namecache_ts {
166 	struct	timespec nc_time;	/* timespec provided by fs */
167 	struct	timespec nc_dotdottime;	/* dotdot timespec provided by fs */
168 	int	nc_ticks;		/* ticks value when entry was added */
169 	int	nc_pad;
170 	struct namecache nc_nc;
171 };
172 
173 TAILQ_HEAD(cache_freebatch, namecache);
174 
175 /*
176  * At least mips n32 performs 64-bit accesses to timespec as found
177  * in namecache_ts and requires them to be aligned. Since others
178  * may be in the same spot suffer a little bit and enforce the
179  * alignment for everyone. Note this is a nop for 64-bit platforms.
180  */
181 #define CACHE_ZONE_ALIGNMENT	UMA_ALIGNOF(time_t)
182 
183 /*
184  * TODO: the initial value of CACHE_PATH_CUTOFF was inherited from the
185  * 4.4 BSD codebase. Later on struct namecache was tweaked to become
186  * smaller and the value was bumped to retain the total size, but it
187  * was never re-evaluated for suitability. A simple test counting
188  * lengths during package building shows that the value of 45 covers
189  * about 86% of all added entries, reaching 99% at 65.
190  *
191  * Regardless of the above, use of dedicated zones instead of malloc may be
192  * inducing additional waste. This may be hard to address as said zones are
193  * tied to VFS SMR. Even if retaining them, the current split should be
194  * re-evaluated.
195  */
196 #ifdef __LP64__
197 #define	CACHE_PATH_CUTOFF	45
198 #define	CACHE_LARGE_PAD		6
199 #else
200 #define	CACHE_PATH_CUTOFF	41
201 #define	CACHE_LARGE_PAD		2
202 #endif
203 
204 #define CACHE_ZONE_SMALL_SIZE		(offsetof(struct namecache, nc_name) + CACHE_PATH_CUTOFF + 1)
205 #define CACHE_ZONE_SMALL_TS_SIZE	(offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_SMALL_SIZE)
206 #define CACHE_ZONE_LARGE_SIZE		(offsetof(struct namecache, nc_name) + NAME_MAX + 1 + CACHE_LARGE_PAD)
207 #define CACHE_ZONE_LARGE_TS_SIZE	(offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_LARGE_SIZE)
208 
209 _Static_assert((CACHE_ZONE_SMALL_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
210 _Static_assert((CACHE_ZONE_SMALL_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
211 _Static_assert((CACHE_ZONE_LARGE_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
212 _Static_assert((CACHE_ZONE_LARGE_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
213 
214 #define	nc_vp		n_un.nu_vp
215 #define	nc_neg		n_un.nu_neg
216 
217 /*
218  * Flags in namecache.nc_flag
219  */
220 #define NCF_WHITE	0x01
221 #define NCF_ISDOTDOT	0x02
222 #define	NCF_TS		0x04
223 #define	NCF_DTS		0x08
224 #define	NCF_DVDROP	0x10
225 #define	NCF_NEGATIVE	0x20
226 #define	NCF_INVALID	0x40
227 #define	NCF_WIP		0x80
228 
229 /*
230  * Flags in negstate.neg_flag
231  */
232 #define NEG_HOT		0x01
233 
234 static bool	cache_neg_evict_cond(u_long lnumcache);
235 
236 /*
237  * Mark an entry as invalid.
238  *
239  * This is called before it starts getting deconstructed.
240  */
241 static void
242 cache_ncp_invalidate(struct namecache *ncp)
243 {
244 
245 	KASSERT((ncp->nc_flag & NCF_INVALID) == 0,
246 	    ("%s: entry %p already invalid", __func__, ncp));
247 	atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_INVALID);
248 	atomic_thread_fence_rel();
249 }
250 
251 /*
252  * Check whether the entry can be safely used.
253  *
254  * All places which elide locks are supposed to call this after they are
255  * done with reading from an entry.
256  */
257 #define cache_ncp_canuse(ncp)	({					\
258 	struct namecache *_ncp = (ncp);					\
259 	u_char _nc_flag;						\
260 									\
261 	atomic_thread_fence_acq();					\
262 	_nc_flag = atomic_load_char(&_ncp->nc_flag);			\
263 	__predict_true((_nc_flag & (NCF_INVALID | NCF_WIP)) == 0);	\
264 })
265 
266 /*
267  * Like the above but also checks NCF_WHITE.
268  */
269 #define cache_fpl_neg_ncp_canuse(ncp)	({				\
270 	struct namecache *_ncp = (ncp);					\
271 	u_char _nc_flag;						\
272 									\
273 	atomic_thread_fence_acq();					\
274 	_nc_flag = atomic_load_char(&_ncp->nc_flag);			\
275 	__predict_true((_nc_flag & (NCF_INVALID | NCF_WIP | NCF_WHITE)) == 0);	\
276 })
277 
278 /*
279  * Name caching works as follows:
280  *
281  * Names found by directory scans are retained in a cache
282  * for future reference.  It is managed LRU, so frequently
283  * used names will hang around.  Cache is indexed by hash value
284  * obtained from (dvp, name) where dvp refers to the directory
285  * containing name.
286  *
287  * If it is a "negative" entry, (i.e. for a name that is known NOT to
288  * exist) the vnode pointer will be NULL.
289  *
290  * Upon reaching the last segment of a path, if the reference
291  * is for DELETE, or NOCACHE is set (rewrite), and the
292  * name is located in the cache, it will be dropped.
293  *
294  * These locks are used (in the order in which they can be taken):
295  * NAME		TYPE	ROLE
296  * vnodelock	mtx	vnode lists and v_cache_dd field protection
297  * bucketlock	mtx	for access to given set of hash buckets
298  * neglist	mtx	negative entry LRU management
299  *
300  * It is legal to take multiple vnodelock and bucketlock locks. The locking
301  * order is lower address first. Both are recursive.
302  *
303  * "." lookups are lockless.
304  *
305  * ".." and vnode -> name lookups require vnodelock.
306  *
307  * name -> vnode lookup requires the relevant bucketlock to be held for reading.
308  *
309  * Insertions and removals of entries require involved vnodes and bucketlocks
310  * to be locked to provide safe operation against other threads modifying the
311  * cache.
312  *
313  * Some lookups result in removal of the found entry (e.g. getting rid of a
314  * negative entry with the intent to create a positive one), which poses a
315  * problem when multiple threads reach the state. Similarly, two different
316  * threads can purge two different vnodes and try to remove the same name.
317  *
318  * If the already held vnode lock is lower than the second required lock, we
319  * can just take the other lock. However, in the opposite case, this could
320  * deadlock. As such, this is resolved by trylocking and if that fails unlocking
321  * the first node, locking everything in order and revalidating the state.
322  */
323 
324 VFS_SMR_DECLARE;
325 
326 static SYSCTL_NODE(_vfs_cache, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
327     "Name cache parameters");
328 
329 static u_int __read_mostly	ncsize; /* the size as computed on creation or resizing */
330 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, size, CTLFLAG_RW, &ncsize, 0,
331     "Total namecache capacity");
332 
333 u_int ncsizefactor = 2;
334 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, sizefactor, CTLFLAG_RW, &ncsizefactor, 0,
335     "Size factor for namecache");
336 
337 static u_long __read_mostly	ncnegfactor = 5; /* ratio of negative entries */
338 SYSCTL_ULONG(_vfs_cache_param, OID_AUTO, negfactor, CTLFLAG_RW, &ncnegfactor, 0,
339     "Ratio of negative namecache entries");
340 
341 /*
342  * Negative entry % of namecache capacity above which automatic eviction is allowed.
343  *
344  * Check cache_neg_evict_cond for details.
345  */
346 static u_int ncnegminpct = 3;
347 
348 static u_int __read_mostly     neg_min; /* the above recomputed against ncsize */
349 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, negmin, CTLFLAG_RD, &neg_min, 0,
350     "Negative entry count above which automatic eviction is allowed");
351 
352 /*
353  * Structures associated with name caching.
354  */
355 #define NCHHASH(hash) \
356 	(&nchashtbl[(hash) & nchash])
357 static __read_mostly CK_SLIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
358 static u_long __read_mostly	nchash;			/* size of hash table */
359 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
360     "Size of namecache hash table");
361 static u_long __exclusive_cache_line	numneg;	/* number of negative entries allocated */
362 static u_long __exclusive_cache_line	numcache;/* number of cache entries allocated */
363 
364 struct nchstats	nchstats;		/* cache effectiveness statistics */
365 
366 static bool __read_frequently cache_fast_revlookup = true;
367 SYSCTL_BOOL(_vfs, OID_AUTO, cache_fast_revlookup, CTLFLAG_RW,
368     &cache_fast_revlookup, 0, "");
369 
370 static u_int __exclusive_cache_line neg_cycle;
371 
372 #define ncneghash	3
373 #define	numneglists	(ncneghash + 1)
374 
375 struct neglist {
376 	struct mtx		nl_evict_lock;
377 	struct mtx		nl_lock __aligned(CACHE_LINE_SIZE);
378 	TAILQ_HEAD(, namecache) nl_list;
379 	TAILQ_HEAD(, namecache) nl_hotlist;
380 	u_long			nl_hotnum;
381 } __aligned(CACHE_LINE_SIZE);
382 
383 static struct neglist neglists[numneglists];
384 
385 static inline struct neglist *
386 NCP2NEGLIST(struct namecache *ncp)
387 {
388 
389 	return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
390 }
391 
392 static inline struct negstate *
393 NCP2NEGSTATE(struct namecache *ncp)
394 {
395 
396 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
397 	return (&ncp->nc_neg);
398 }
399 
400 #define	numbucketlocks (ncbuckethash + 1)
401 static u_int __read_mostly  ncbuckethash;
402 static struct mtx_padalign __read_mostly  *bucketlocks;
403 #define	HASH2BUCKETLOCK(hash) \
404 	((struct mtx *)(&bucketlocks[((hash) & ncbuckethash)]))
405 
406 #define	numvnodelocks (ncvnodehash + 1)
407 static u_int __read_mostly  ncvnodehash;
408 static struct mtx __read_mostly *vnodelocks;
409 static inline struct mtx *
410 VP2VNODELOCK(struct vnode *vp)
411 {
412 
413 	return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
414 }
415 
416 static void
417 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
418 {
419 	struct namecache_ts *ncp_ts;
420 
421 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
422 	    (tsp == NULL && ticksp == NULL),
423 	    ("No NCF_TS"));
424 
425 	if (tsp == NULL)
426 		return;
427 
428 	ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
429 	*tsp = ncp_ts->nc_time;
430 	*ticksp = ncp_ts->nc_ticks;
431 }
432 
433 #ifdef DEBUG_CACHE
434 static int __read_mostly	doingcache = 1;	/* 1 => enable the cache */
435 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
436     "VFS namecache enabled");
437 #endif
438 
439 /* Export size information to userland */
440 SYSCTL_INT(_debug_sizeof, OID_AUTO, namecache, CTLFLAG_RD, SYSCTL_NULL_INT_PTR,
441     sizeof(struct namecache), "sizeof(struct namecache)");
442 
443 /*
444  * The new name cache statistics
445  */
446 static SYSCTL_NODE(_vfs_cache, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
447     "Name cache statistics");
448 
449 #define STATNODE_ULONG(name, varname, descr)					\
450 	SYSCTL_ULONG(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
451 #define STATNODE_COUNTER(name, varname, descr)					\
452 	static COUNTER_U64_DEFINE_EARLY(varname);				\
453 	SYSCTL_COUNTER_U64(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, \
454 	    descr);
455 STATNODE_ULONG(neg, numneg, "Number of negative cache entries");
456 STATNODE_ULONG(count, numcache, "Number of cache entries");
457 STATNODE_COUNTER(heldvnodes, numcachehv, "Number of namecache entries with vnodes held");
458 STATNODE_COUNTER(drops, numdrops, "Number of dropped entries due to reaching the limit");
459 STATNODE_COUNTER(dothits, dothits, "Number of '.' hits");
460 STATNODE_COUNTER(dotdothis, dotdothits, "Number of '..' hits");
461 STATNODE_COUNTER(miss, nummiss, "Number of cache misses");
462 STATNODE_COUNTER(misszap, nummisszap, "Number of cache misses we do not want to cache");
463 STATNODE_COUNTER(posszaps, numposzaps,
464     "Number of cache hits (positive) we do not want to cache");
465 STATNODE_COUNTER(poshits, numposhits, "Number of cache hits (positive)");
466 STATNODE_COUNTER(negzaps, numnegzaps,
467     "Number of cache hits (negative) we do not want to cache");
468 STATNODE_COUNTER(neghits, numneghits, "Number of cache hits (negative)");
469 /* These count for vn_getcwd(), too. */
470 STATNODE_COUNTER(fullpathcalls, numfullpathcalls, "Number of fullpath search calls");
471 STATNODE_COUNTER(fullpathfail1, numfullpathfail1, "Number of fullpath search errors (ENOTDIR)");
472 STATNODE_COUNTER(fullpathfail2, numfullpathfail2,
473     "Number of fullpath search errors (VOP_VPTOCNP failures)");
474 STATNODE_COUNTER(fullpathfail4, numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
475 STATNODE_COUNTER(fullpathfound, numfullpathfound, "Number of successful fullpath calls");
476 
477 /*
478  * Debug or developer statistics.
479  */
480 static SYSCTL_NODE(_vfs_cache, OID_AUTO, debug, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
481     "Name cache debugging");
482 #define DEBUGNODE_ULONG(name, varname, descr)					\
483 	SYSCTL_ULONG(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
484 #define DEBUGNODE_COUNTER(name, varname, descr)					\
485 	static COUNTER_U64_DEFINE_EARLY(varname);				\
486 	SYSCTL_COUNTER_U64(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, \
487 	    descr);
488 DEBUGNODE_COUNTER(zap_bucket_relock_success, zap_bucket_relock_success,
489     "Number of successful removals after relocking");
490 static long zap_bucket_fail;
491 DEBUGNODE_ULONG(zap_bucket_fail, zap_bucket_fail, "");
492 static long zap_bucket_fail2;
493 DEBUGNODE_ULONG(zap_bucket_fail2, zap_bucket_fail2, "");
494 static long cache_lock_vnodes_cel_3_failures;
495 DEBUGNODE_ULONG(vnodes_cel_3_failures, cache_lock_vnodes_cel_3_failures,
496     "Number of times 3-way vnode locking failed");
497 
498 static void cache_zap_locked(struct namecache *ncp);
499 static int vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf,
500     char **freebuf, size_t *buflen);
501 static int vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
502     char **retbuf, size_t *buflen, size_t addend);
503 static int vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf,
504     char **retbuf, size_t *buflen);
505 static int vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf,
506     char **retbuf, size_t *len, size_t addend);
507 
508 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
509 
510 static inline void
511 cache_assert_vlp_locked(struct mtx *vlp)
512 {
513 
514 	if (vlp != NULL)
515 		mtx_assert(vlp, MA_OWNED);
516 }
517 
518 static inline void
519 cache_assert_vnode_locked(struct vnode *vp)
520 {
521 	struct mtx *vlp;
522 
523 	vlp = VP2VNODELOCK(vp);
524 	cache_assert_vlp_locked(vlp);
525 }
526 
527 /*
528  * Directory vnodes with entries are held for two reasons:
529  * 1. make them less of a target for reclamation in vnlru
530  * 2. suffer smaller performance penalty in locked lookup as requeieing is avoided
531  *
532  * It will be feasible to stop doing it altogether if all filesystems start
533  * supporting lockless lookup.
534  */
535 static void
536 cache_hold_vnode(struct vnode *vp)
537 {
538 
539 	cache_assert_vnode_locked(vp);
540 	VNPASS(LIST_EMPTY(&vp->v_cache_src), vp);
541 	vhold(vp);
542 	counter_u64_add(numcachehv, 1);
543 }
544 
545 static void
546 cache_drop_vnode(struct vnode *vp)
547 {
548 
549 	/*
550 	 * Called after all locks are dropped, meaning we can't assert
551 	 * on the state of v_cache_src.
552 	 */
553 	vdrop(vp);
554 	counter_u64_add(numcachehv, -1);
555 }
556 
557 /*
558  * UMA zones.
559  */
560 static uma_zone_t __read_mostly cache_zone_small;
561 static uma_zone_t __read_mostly cache_zone_small_ts;
562 static uma_zone_t __read_mostly cache_zone_large;
563 static uma_zone_t __read_mostly cache_zone_large_ts;
564 
565 static struct namecache *
566 cache_alloc_uma(int len, bool ts)
567 {
568 	struct namecache_ts *ncp_ts;
569 	struct namecache *ncp;
570 
571 	if (__predict_false(ts)) {
572 		if (len <= CACHE_PATH_CUTOFF)
573 			ncp_ts = uma_zalloc_smr(cache_zone_small_ts, M_WAITOK);
574 		else
575 			ncp_ts = uma_zalloc_smr(cache_zone_large_ts, M_WAITOK);
576 		ncp = &ncp_ts->nc_nc;
577 	} else {
578 		if (len <= CACHE_PATH_CUTOFF)
579 			ncp = uma_zalloc_smr(cache_zone_small, M_WAITOK);
580 		else
581 			ncp = uma_zalloc_smr(cache_zone_large, M_WAITOK);
582 	}
583 	return (ncp);
584 }
585 
586 static void
587 cache_free_uma(struct namecache *ncp)
588 {
589 	struct namecache_ts *ncp_ts;
590 
591 	if (__predict_false(ncp->nc_flag & NCF_TS)) {
592 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
593 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
594 			uma_zfree_smr(cache_zone_small_ts, ncp_ts);
595 		else
596 			uma_zfree_smr(cache_zone_large_ts, ncp_ts);
597 	} else {
598 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
599 			uma_zfree_smr(cache_zone_small, ncp);
600 		else
601 			uma_zfree_smr(cache_zone_large, ncp);
602 	}
603 }
604 
605 static struct namecache *
606 cache_alloc(int len, bool ts)
607 {
608 	u_long lnumcache;
609 
610 	/*
611 	 * Avoid blowout in namecache entries.
612 	 *
613 	 * Bugs:
614 	 * 1. filesystems may end up trying to add an already existing entry
615 	 * (for example this can happen after a cache miss during concurrent
616 	 * lookup), in which case we will call cache_neg_evict despite not
617 	 * adding anything.
618 	 * 2. the routine may fail to free anything and no provisions are made
619 	 * to make it try harder (see the inside for failure modes)
620 	 * 3. it only ever looks at negative entries.
621 	 */
622 	lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
623 	if (cache_neg_evict_cond(lnumcache)) {
624 		lnumcache = atomic_load_long(&numcache);
625 	}
626 	if (__predict_false(lnumcache >= ncsize)) {
627 		atomic_subtract_long(&numcache, 1);
628 		counter_u64_add(numdrops, 1);
629 		return (NULL);
630 	}
631 	return (cache_alloc_uma(len, ts));
632 }
633 
634 static void
635 cache_free(struct namecache *ncp)
636 {
637 
638 	MPASS(ncp != NULL);
639 	if ((ncp->nc_flag & NCF_DVDROP) != 0) {
640 		cache_drop_vnode(ncp->nc_dvp);
641 	}
642 	cache_free_uma(ncp);
643 	atomic_subtract_long(&numcache, 1);
644 }
645 
646 static void
647 cache_free_batch(struct cache_freebatch *batch)
648 {
649 	struct namecache *ncp, *nnp;
650 	int i;
651 
652 	i = 0;
653 	if (TAILQ_EMPTY(batch))
654 		goto out;
655 	TAILQ_FOREACH_SAFE(ncp, batch, nc_dst, nnp) {
656 		if ((ncp->nc_flag & NCF_DVDROP) != 0) {
657 			cache_drop_vnode(ncp->nc_dvp);
658 		}
659 		cache_free_uma(ncp);
660 		i++;
661 	}
662 	atomic_subtract_long(&numcache, i);
663 out:
664 	SDT_PROBE1(vfs, namecache, purge, batch, i);
665 }
666 
667 /*
668  * TODO: With the value stored we can do better than computing the hash based
669  * on the address. The choice of FNV should also be revisited.
670  */
671 static void
672 cache_prehash(struct vnode *vp)
673 {
674 
675 	vp->v_nchash = fnv_32_buf(&vp, sizeof(vp), FNV1_32_INIT);
676 }
677 
678 static uint32_t
679 cache_get_hash(char *name, u_char len, struct vnode *dvp)
680 {
681 
682 	return (fnv_32_buf(name, len, dvp->v_nchash));
683 }
684 
685 static inline struct nchashhead *
686 NCP2BUCKET(struct namecache *ncp)
687 {
688 	uint32_t hash;
689 
690 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
691 	return (NCHHASH(hash));
692 }
693 
694 static inline struct mtx *
695 NCP2BUCKETLOCK(struct namecache *ncp)
696 {
697 	uint32_t hash;
698 
699 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
700 	return (HASH2BUCKETLOCK(hash));
701 }
702 
703 #ifdef INVARIANTS
704 static void
705 cache_assert_bucket_locked(struct namecache *ncp)
706 {
707 	struct mtx *blp;
708 
709 	blp = NCP2BUCKETLOCK(ncp);
710 	mtx_assert(blp, MA_OWNED);
711 }
712 
713 static void
714 cache_assert_bucket_unlocked(struct namecache *ncp)
715 {
716 	struct mtx *blp;
717 
718 	blp = NCP2BUCKETLOCK(ncp);
719 	mtx_assert(blp, MA_NOTOWNED);
720 }
721 #else
722 #define cache_assert_bucket_locked(x) do { } while (0)
723 #define cache_assert_bucket_unlocked(x) do { } while (0)
724 #endif
725 
726 #define cache_sort_vnodes(x, y)	_cache_sort_vnodes((void **)(x), (void **)(y))
727 static void
728 _cache_sort_vnodes(void **p1, void **p2)
729 {
730 	void *tmp;
731 
732 	MPASS(*p1 != NULL || *p2 != NULL);
733 
734 	if (*p1 > *p2) {
735 		tmp = *p2;
736 		*p2 = *p1;
737 		*p1 = tmp;
738 	}
739 }
740 
741 static void
742 cache_lock_all_buckets(void)
743 {
744 	u_int i;
745 
746 	for (i = 0; i < numbucketlocks; i++)
747 		mtx_lock(&bucketlocks[i]);
748 }
749 
750 static void
751 cache_unlock_all_buckets(void)
752 {
753 	u_int i;
754 
755 	for (i = 0; i < numbucketlocks; i++)
756 		mtx_unlock(&bucketlocks[i]);
757 }
758 
759 static void
760 cache_lock_all_vnodes(void)
761 {
762 	u_int i;
763 
764 	for (i = 0; i < numvnodelocks; i++)
765 		mtx_lock(&vnodelocks[i]);
766 }
767 
768 static void
769 cache_unlock_all_vnodes(void)
770 {
771 	u_int i;
772 
773 	for (i = 0; i < numvnodelocks; i++)
774 		mtx_unlock(&vnodelocks[i]);
775 }
776 
777 static int
778 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
779 {
780 
781 	cache_sort_vnodes(&vlp1, &vlp2);
782 
783 	if (vlp1 != NULL) {
784 		if (!mtx_trylock(vlp1))
785 			return (EAGAIN);
786 	}
787 	if (!mtx_trylock(vlp2)) {
788 		if (vlp1 != NULL)
789 			mtx_unlock(vlp1);
790 		return (EAGAIN);
791 	}
792 
793 	return (0);
794 }
795 
796 static void
797 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
798 {
799 
800 	MPASS(vlp1 != NULL || vlp2 != NULL);
801 	MPASS(vlp1 <= vlp2);
802 
803 	if (vlp1 != NULL)
804 		mtx_lock(vlp1);
805 	if (vlp2 != NULL)
806 		mtx_lock(vlp2);
807 }
808 
809 static void
810 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
811 {
812 
813 	MPASS(vlp1 != NULL || vlp2 != NULL);
814 
815 	if (vlp1 != NULL)
816 		mtx_unlock(vlp1);
817 	if (vlp2 != NULL)
818 		mtx_unlock(vlp2);
819 }
820 
821 static int
822 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
823 {
824 	struct nchstats snap;
825 
826 	if (req->oldptr == NULL)
827 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
828 
829 	snap = nchstats;
830 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
831 	snap.ncs_neghits = counter_u64_fetch(numneghits);
832 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
833 	    counter_u64_fetch(numnegzaps);
834 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
835 	    counter_u64_fetch(nummiss);
836 
837 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
838 }
839 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
840     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
841     "VFS cache effectiveness statistics");
842 
843 static void
844 cache_recalc_neg_min(u_int val)
845 {
846 
847 	neg_min = (ncsize * val) / 100;
848 }
849 
850 static int
851 sysctl_negminpct(SYSCTL_HANDLER_ARGS)
852 {
853 	u_int val;
854 	int error;
855 
856 	val = ncnegminpct;
857 	error = sysctl_handle_int(oidp, &val, 0, req);
858 	if (error != 0 || req->newptr == NULL)
859 		return (error);
860 
861 	if (val == ncnegminpct)
862 		return (0);
863 	if (val < 0 || val > 99)
864 		return (EINVAL);
865 	ncnegminpct = val;
866 	cache_recalc_neg_min(val);
867 	return (0);
868 }
869 
870 SYSCTL_PROC(_vfs_cache_param, OID_AUTO, negminpct,
871     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_negminpct,
872     "I", "Negative entry \% of namecache capacity above which automatic eviction is allowed");
873 
874 #ifdef DIAGNOSTIC
875 /*
876  * Grab an atomic snapshot of the name cache hash chain lengths
877  */
878 static SYSCTL_NODE(_debug, OID_AUTO, hashstat,
879     CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
880     "hash table stats");
881 
882 static int
883 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
884 {
885 	struct nchashhead *ncpp;
886 	struct namecache *ncp;
887 	int i, error, n_nchash, *cntbuf;
888 
889 retry:
890 	n_nchash = nchash + 1;	/* nchash is max index, not count */
891 	if (req->oldptr == NULL)
892 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
893 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
894 	cache_lock_all_buckets();
895 	if (n_nchash != nchash + 1) {
896 		cache_unlock_all_buckets();
897 		free(cntbuf, M_TEMP);
898 		goto retry;
899 	}
900 	/* Scan hash tables counting entries */
901 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
902 		CK_SLIST_FOREACH(ncp, ncpp, nc_hash)
903 			cntbuf[i]++;
904 	cache_unlock_all_buckets();
905 	for (error = 0, i = 0; i < n_nchash; i++)
906 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
907 			break;
908 	free(cntbuf, M_TEMP);
909 	return (error);
910 }
911 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
912     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
913     "nchash chain lengths");
914 
915 static int
916 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
917 {
918 	int error;
919 	struct nchashhead *ncpp;
920 	struct namecache *ncp;
921 	int n_nchash;
922 	int count, maxlength, used, pct;
923 
924 	if (!req->oldptr)
925 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
926 
927 	cache_lock_all_buckets();
928 	n_nchash = nchash + 1;	/* nchash is max index, not count */
929 	used = 0;
930 	maxlength = 0;
931 
932 	/* Scan hash tables for applicable entries */
933 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
934 		count = 0;
935 		CK_SLIST_FOREACH(ncp, ncpp, nc_hash) {
936 			count++;
937 		}
938 		if (count)
939 			used++;
940 		if (maxlength < count)
941 			maxlength = count;
942 	}
943 	n_nchash = nchash + 1;
944 	cache_unlock_all_buckets();
945 	pct = (used * 100) / (n_nchash / 100);
946 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
947 	if (error)
948 		return (error);
949 	error = SYSCTL_OUT(req, &used, sizeof(used));
950 	if (error)
951 		return (error);
952 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
953 	if (error)
954 		return (error);
955 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
956 	if (error)
957 		return (error);
958 	return (0);
959 }
960 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
961     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
962     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
963 #endif
964 
965 /*
966  * Negative entries management
967  *
968  * Various workloads create plenty of negative entries and barely use them
969  * afterwards. Moreover malicious users can keep performing bogus lookups
970  * adding even more entries. For example "make tinderbox" as of writing this
971  * comment ends up with 2.6M namecache entries in total, 1.2M of which are
972  * negative.
973  *
974  * As such, a rather aggressive eviction method is needed. The currently
975  * employed method is a placeholder.
976  *
977  * Entries are split over numneglists separate lists, each of which is further
978  * split into hot and cold entries. Entries get promoted after getting a hit.
979  * Eviction happens on addition of new entry.
980  */
981 static SYSCTL_NODE(_vfs_cache, OID_AUTO, neg, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
982     "Name cache negative entry statistics");
983 
984 SYSCTL_ULONG(_vfs_cache_neg, OID_AUTO, count, CTLFLAG_RD, &numneg, 0,
985     "Number of negative cache entries");
986 
987 static COUNTER_U64_DEFINE_EARLY(neg_created);
988 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, created, CTLFLAG_RD, &neg_created,
989     "Number of created negative entries");
990 
991 static COUNTER_U64_DEFINE_EARLY(neg_evicted);
992 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evicted, CTLFLAG_RD, &neg_evicted,
993     "Number of evicted negative entries");
994 
995 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_empty);
996 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_empty, CTLFLAG_RD,
997     &neg_evict_skipped_empty,
998     "Number of times evicting failed due to lack of entries");
999 
1000 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_missed);
1001 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_missed, CTLFLAG_RD,
1002     &neg_evict_skipped_missed,
1003     "Number of times evicting failed due to target entry disappearing");
1004 
1005 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_contended);
1006 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_contended, CTLFLAG_RD,
1007     &neg_evict_skipped_contended,
1008     "Number of times evicting failed due to contention");
1009 
1010 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, hits, CTLFLAG_RD, &numneghits,
1011     "Number of cache hits (negative)");
1012 
1013 static int
1014 sysctl_neg_hot(SYSCTL_HANDLER_ARGS)
1015 {
1016 	int i, out;
1017 
1018 	out = 0;
1019 	for (i = 0; i < numneglists; i++)
1020 		out += neglists[i].nl_hotnum;
1021 
1022 	return (SYSCTL_OUT(req, &out, sizeof(out)));
1023 }
1024 SYSCTL_PROC(_vfs_cache_neg, OID_AUTO, hot, CTLTYPE_INT | CTLFLAG_RD |
1025     CTLFLAG_MPSAFE, 0, 0, sysctl_neg_hot, "I",
1026     "Number of hot negative entries");
1027 
1028 static void
1029 cache_neg_init(struct namecache *ncp)
1030 {
1031 	struct negstate *ns;
1032 
1033 	ncp->nc_flag |= NCF_NEGATIVE;
1034 	ns = NCP2NEGSTATE(ncp);
1035 	ns->neg_flag = 0;
1036 	ns->neg_hit = 0;
1037 	counter_u64_add(neg_created, 1);
1038 }
1039 
1040 #define CACHE_NEG_PROMOTION_THRESH 2
1041 
1042 static bool
1043 cache_neg_hit_prep(struct namecache *ncp)
1044 {
1045 	struct negstate *ns;
1046 	u_char n;
1047 
1048 	ns = NCP2NEGSTATE(ncp);
1049 	n = atomic_load_char(&ns->neg_hit);
1050 	for (;;) {
1051 		if (n >= CACHE_NEG_PROMOTION_THRESH)
1052 			return (false);
1053 		if (atomic_fcmpset_8(&ns->neg_hit, &n, n + 1))
1054 			break;
1055 	}
1056 	return (n + 1 == CACHE_NEG_PROMOTION_THRESH);
1057 }
1058 
1059 /*
1060  * Nothing to do here but it is provided for completeness as some
1061  * cache_neg_hit_prep callers may end up returning without even
1062  * trying to promote.
1063  */
1064 #define cache_neg_hit_abort(ncp)	do { } while (0)
1065 
1066 static void
1067 cache_neg_hit_finish(struct namecache *ncp)
1068 {
1069 
1070 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, ncp->nc_dvp, ncp->nc_name);
1071 	counter_u64_add(numneghits, 1);
1072 }
1073 
1074 /*
1075  * Move a negative entry to the hot list.
1076  */
1077 static void
1078 cache_neg_promote_locked(struct namecache *ncp)
1079 {
1080 	struct neglist *nl;
1081 	struct negstate *ns;
1082 
1083 	ns = NCP2NEGSTATE(ncp);
1084 	nl = NCP2NEGLIST(ncp);
1085 	mtx_assert(&nl->nl_lock, MA_OWNED);
1086 	if ((ns->neg_flag & NEG_HOT) == 0) {
1087 		TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1088 		TAILQ_INSERT_TAIL(&nl->nl_hotlist, ncp, nc_dst);
1089 		nl->nl_hotnum++;
1090 		ns->neg_flag |= NEG_HOT;
1091 	}
1092 }
1093 
1094 /*
1095  * Move a hot negative entry to the cold list.
1096  */
1097 static void
1098 cache_neg_demote_locked(struct namecache *ncp)
1099 {
1100 	struct neglist *nl;
1101 	struct negstate *ns;
1102 
1103 	ns = NCP2NEGSTATE(ncp);
1104 	nl = NCP2NEGLIST(ncp);
1105 	mtx_assert(&nl->nl_lock, MA_OWNED);
1106 	MPASS(ns->neg_flag & NEG_HOT);
1107 	TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1108 	TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1109 	nl->nl_hotnum--;
1110 	ns->neg_flag &= ~NEG_HOT;
1111 	atomic_store_char(&ns->neg_hit, 0);
1112 }
1113 
1114 /*
1115  * Move a negative entry to the hot list if it matches the lookup.
1116  *
1117  * We have to take locks, but they may be contended and in the worst
1118  * case we may need to go off CPU. We don't want to spin within the
1119  * smr section and we can't block with it. Exiting the section means
1120  * the found entry could have been evicted. We are going to look it
1121  * up again.
1122  */
1123 static bool
1124 cache_neg_promote_cond(struct vnode *dvp, struct componentname *cnp,
1125     struct namecache *oncp, uint32_t hash)
1126 {
1127 	struct namecache *ncp;
1128 	struct neglist *nl;
1129 	u_char nc_flag;
1130 
1131 	nl = NCP2NEGLIST(oncp);
1132 
1133 	mtx_lock(&nl->nl_lock);
1134 	/*
1135 	 * For hash iteration.
1136 	 */
1137 	vfs_smr_enter();
1138 
1139 	/*
1140 	 * Avoid all surprises by only succeeding if we got the same entry and
1141 	 * bailing completely otherwise.
1142 	 * XXX There are no provisions to keep the vnode around, meaning we may
1143 	 * end up promoting a negative entry for a *new* vnode and returning
1144 	 * ENOENT on its account. This is the error we want to return anyway
1145 	 * and promotion is harmless.
1146 	 *
1147 	 * In particular at this point there can be a new ncp which matches the
1148 	 * search but hashes to a different neglist.
1149 	 */
1150 	CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1151 		if (ncp == oncp)
1152 			break;
1153 	}
1154 
1155 	/*
1156 	 * No match to begin with.
1157 	 */
1158 	if (__predict_false(ncp == NULL)) {
1159 		goto out_abort;
1160 	}
1161 
1162 	/*
1163 	 * The newly found entry may be something different...
1164 	 */
1165 	if (!(ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1166 	    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))) {
1167 		goto out_abort;
1168 	}
1169 
1170 	/*
1171 	 * ... and not even negative.
1172 	 */
1173 	nc_flag = atomic_load_char(&ncp->nc_flag);
1174 	if ((nc_flag & NCF_NEGATIVE) == 0) {
1175 		goto out_abort;
1176 	}
1177 
1178 	if (!cache_ncp_canuse(ncp)) {
1179 		goto out_abort;
1180 	}
1181 
1182 	cache_neg_promote_locked(ncp);
1183 	cache_neg_hit_finish(ncp);
1184 	vfs_smr_exit();
1185 	mtx_unlock(&nl->nl_lock);
1186 	return (true);
1187 out_abort:
1188 	vfs_smr_exit();
1189 	mtx_unlock(&nl->nl_lock);
1190 	return (false);
1191 }
1192 
1193 static void
1194 cache_neg_promote(struct namecache *ncp)
1195 {
1196 	struct neglist *nl;
1197 
1198 	nl = NCP2NEGLIST(ncp);
1199 	mtx_lock(&nl->nl_lock);
1200 	cache_neg_promote_locked(ncp);
1201 	mtx_unlock(&nl->nl_lock);
1202 }
1203 
1204 static void
1205 cache_neg_insert(struct namecache *ncp)
1206 {
1207 	struct neglist *nl;
1208 
1209 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
1210 	cache_assert_bucket_locked(ncp);
1211 	nl = NCP2NEGLIST(ncp);
1212 	mtx_lock(&nl->nl_lock);
1213 	TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1214 	mtx_unlock(&nl->nl_lock);
1215 	atomic_add_long(&numneg, 1);
1216 }
1217 
1218 static void
1219 cache_neg_remove(struct namecache *ncp)
1220 {
1221 	struct neglist *nl;
1222 	struct negstate *ns;
1223 
1224 	cache_assert_bucket_locked(ncp);
1225 	nl = NCP2NEGLIST(ncp);
1226 	ns = NCP2NEGSTATE(ncp);
1227 	mtx_lock(&nl->nl_lock);
1228 	if ((ns->neg_flag & NEG_HOT) != 0) {
1229 		TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1230 		nl->nl_hotnum--;
1231 	} else {
1232 		TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1233 	}
1234 	mtx_unlock(&nl->nl_lock);
1235 	atomic_subtract_long(&numneg, 1);
1236 }
1237 
1238 static struct neglist *
1239 cache_neg_evict_select_list(void)
1240 {
1241 	struct neglist *nl;
1242 	u_int c;
1243 
1244 	c = atomic_fetchadd_int(&neg_cycle, 1) + 1;
1245 	nl = &neglists[c % numneglists];
1246 	if (!mtx_trylock(&nl->nl_evict_lock)) {
1247 		counter_u64_add(neg_evict_skipped_contended, 1);
1248 		return (NULL);
1249 	}
1250 	return (nl);
1251 }
1252 
1253 static struct namecache *
1254 cache_neg_evict_select_entry(struct neglist *nl)
1255 {
1256 	struct namecache *ncp, *lncp;
1257 	struct negstate *ns, *lns;
1258 	int i;
1259 
1260 	mtx_assert(&nl->nl_evict_lock, MA_OWNED);
1261 	mtx_assert(&nl->nl_lock, MA_OWNED);
1262 	ncp = TAILQ_FIRST(&nl->nl_list);
1263 	if (ncp == NULL)
1264 		return (NULL);
1265 	lncp = ncp;
1266 	lns = NCP2NEGSTATE(lncp);
1267 	for (i = 1; i < 4; i++) {
1268 		ncp = TAILQ_NEXT(ncp, nc_dst);
1269 		if (ncp == NULL)
1270 			break;
1271 		ns = NCP2NEGSTATE(ncp);
1272 		if (ns->neg_hit < lns->neg_hit) {
1273 			lncp = ncp;
1274 			lns = ns;
1275 		}
1276 	}
1277 	return (lncp);
1278 }
1279 
1280 static bool
1281 cache_neg_evict(void)
1282 {
1283 	struct namecache *ncp, *ncp2;
1284 	struct neglist *nl;
1285 	struct vnode *dvp;
1286 	struct mtx *dvlp;
1287 	struct mtx *blp;
1288 	uint32_t hash;
1289 	u_char nlen;
1290 	bool evicted;
1291 
1292 	nl = cache_neg_evict_select_list();
1293 	if (nl == NULL) {
1294 		return (false);
1295 	}
1296 
1297 	mtx_lock(&nl->nl_lock);
1298 	ncp = TAILQ_FIRST(&nl->nl_hotlist);
1299 	if (ncp != NULL) {
1300 		cache_neg_demote_locked(ncp);
1301 	}
1302 	ncp = cache_neg_evict_select_entry(nl);
1303 	if (ncp == NULL) {
1304 		counter_u64_add(neg_evict_skipped_empty, 1);
1305 		mtx_unlock(&nl->nl_lock);
1306 		mtx_unlock(&nl->nl_evict_lock);
1307 		return (false);
1308 	}
1309 	nlen = ncp->nc_nlen;
1310 	dvp = ncp->nc_dvp;
1311 	hash = cache_get_hash(ncp->nc_name, nlen, dvp);
1312 	dvlp = VP2VNODELOCK(dvp);
1313 	blp = HASH2BUCKETLOCK(hash);
1314 	mtx_unlock(&nl->nl_lock);
1315 	mtx_unlock(&nl->nl_evict_lock);
1316 	mtx_lock(dvlp);
1317 	mtx_lock(blp);
1318 	/*
1319 	 * Note that since all locks were dropped above, the entry may be
1320 	 * gone or reallocated to be something else.
1321 	 */
1322 	CK_SLIST_FOREACH(ncp2, (NCHHASH(hash)), nc_hash) {
1323 		if (ncp2 == ncp && ncp2->nc_dvp == dvp &&
1324 		    ncp2->nc_nlen == nlen && (ncp2->nc_flag & NCF_NEGATIVE) != 0)
1325 			break;
1326 	}
1327 	if (ncp2 == NULL) {
1328 		counter_u64_add(neg_evict_skipped_missed, 1);
1329 		ncp = NULL;
1330 		evicted = false;
1331 	} else {
1332 		MPASS(dvlp == VP2VNODELOCK(ncp->nc_dvp));
1333 		MPASS(blp == NCP2BUCKETLOCK(ncp));
1334 		SDT_PROBE2(vfs, namecache, evict_negative, done, ncp->nc_dvp,
1335 		    ncp->nc_name);
1336 		cache_zap_locked(ncp);
1337 		counter_u64_add(neg_evicted, 1);
1338 		evicted = true;
1339 	}
1340 	mtx_unlock(blp);
1341 	mtx_unlock(dvlp);
1342 	if (ncp != NULL)
1343 		cache_free(ncp);
1344 	return (evicted);
1345 }
1346 
1347 /*
1348  * Maybe evict a negative entry to create more room.
1349  *
1350  * The ncnegfactor parameter limits what fraction of the total count
1351  * can comprise of negative entries. However, if the cache is just
1352  * warming up this leads to excessive evictions.  As such, ncnegminpct
1353  * (recomputed to neg_min) dictates whether the above should be
1354  * applied.
1355  *
1356  * Try evicting if the cache is close to full capacity regardless of
1357  * other considerations.
1358  */
1359 static bool
1360 cache_neg_evict_cond(u_long lnumcache)
1361 {
1362 	u_long lnumneg;
1363 
1364 	if (ncsize - 1000 < lnumcache)
1365 		goto out_evict;
1366 	lnumneg = atomic_load_long(&numneg);
1367 	if (lnumneg < neg_min)
1368 		return (false);
1369 	if (lnumneg * ncnegfactor < lnumcache)
1370 		return (false);
1371 out_evict:
1372 	return (cache_neg_evict());
1373 }
1374 
1375 /*
1376  * cache_zap_locked():
1377  *
1378  *   Removes a namecache entry from cache, whether it contains an actual
1379  *   pointer to a vnode or if it is just a negative cache entry.
1380  */
1381 static void
1382 cache_zap_locked(struct namecache *ncp)
1383 {
1384 	struct nchashhead *ncpp;
1385 
1386 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1387 		cache_assert_vnode_locked(ncp->nc_vp);
1388 	cache_assert_vnode_locked(ncp->nc_dvp);
1389 	cache_assert_bucket_locked(ncp);
1390 
1391 	cache_ncp_invalidate(ncp);
1392 
1393 	ncpp = NCP2BUCKET(ncp);
1394 	CK_SLIST_REMOVE(ncpp, ncp, namecache, nc_hash);
1395 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1396 		SDT_PROBE3(vfs, namecache, zap, done, ncp->nc_dvp,
1397 		    ncp->nc_name, ncp->nc_vp);
1398 		TAILQ_REMOVE(&ncp->nc_vp->v_cache_dst, ncp, nc_dst);
1399 		if (ncp == ncp->nc_vp->v_cache_dd) {
1400 			vn_seqc_write_begin_unheld(ncp->nc_vp);
1401 			ncp->nc_vp->v_cache_dd = NULL;
1402 			vn_seqc_write_end(ncp->nc_vp);
1403 		}
1404 	} else {
1405 		SDT_PROBE2(vfs, namecache, zap_negative, done, ncp->nc_dvp,
1406 		    ncp->nc_name);
1407 		cache_neg_remove(ncp);
1408 	}
1409 	if (ncp->nc_flag & NCF_ISDOTDOT) {
1410 		if (ncp == ncp->nc_dvp->v_cache_dd) {
1411 			vn_seqc_write_begin_unheld(ncp->nc_dvp);
1412 			ncp->nc_dvp->v_cache_dd = NULL;
1413 			vn_seqc_write_end(ncp->nc_dvp);
1414 		}
1415 	} else {
1416 		LIST_REMOVE(ncp, nc_src);
1417 		if (LIST_EMPTY(&ncp->nc_dvp->v_cache_src)) {
1418 			ncp->nc_flag |= NCF_DVDROP;
1419 		}
1420 	}
1421 }
1422 
1423 static void
1424 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
1425 {
1426 	struct mtx *blp;
1427 
1428 	MPASS(ncp->nc_dvp == vp);
1429 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
1430 	cache_assert_vnode_locked(vp);
1431 
1432 	blp = NCP2BUCKETLOCK(ncp);
1433 	mtx_lock(blp);
1434 	cache_zap_locked(ncp);
1435 	mtx_unlock(blp);
1436 }
1437 
1438 static bool
1439 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
1440     struct mtx **vlpp)
1441 {
1442 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
1443 	struct mtx *blp;
1444 
1445 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
1446 	cache_assert_vnode_locked(vp);
1447 
1448 	if (ncp->nc_flag & NCF_NEGATIVE) {
1449 		if (*vlpp != NULL) {
1450 			mtx_unlock(*vlpp);
1451 			*vlpp = NULL;
1452 		}
1453 		cache_zap_negative_locked_vnode_kl(ncp, vp);
1454 		return (true);
1455 	}
1456 
1457 	pvlp = VP2VNODELOCK(vp);
1458 	blp = NCP2BUCKETLOCK(ncp);
1459 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
1460 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
1461 
1462 	if (*vlpp == vlp1 || *vlpp == vlp2) {
1463 		to_unlock = *vlpp;
1464 		*vlpp = NULL;
1465 	} else {
1466 		if (*vlpp != NULL) {
1467 			mtx_unlock(*vlpp);
1468 			*vlpp = NULL;
1469 		}
1470 		cache_sort_vnodes(&vlp1, &vlp2);
1471 		if (vlp1 == pvlp) {
1472 			mtx_lock(vlp2);
1473 			to_unlock = vlp2;
1474 		} else {
1475 			if (!mtx_trylock(vlp1))
1476 				goto out_relock;
1477 			to_unlock = vlp1;
1478 		}
1479 	}
1480 	mtx_lock(blp);
1481 	cache_zap_locked(ncp);
1482 	mtx_unlock(blp);
1483 	if (to_unlock != NULL)
1484 		mtx_unlock(to_unlock);
1485 	return (true);
1486 
1487 out_relock:
1488 	mtx_unlock(vlp2);
1489 	mtx_lock(vlp1);
1490 	mtx_lock(vlp2);
1491 	MPASS(*vlpp == NULL);
1492 	*vlpp = vlp1;
1493 	return (false);
1494 }
1495 
1496 /*
1497  * If trylocking failed we can get here. We know enough to take all needed locks
1498  * in the right order and re-lookup the entry.
1499  */
1500 static int
1501 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1502     struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1503     struct mtx *blp)
1504 {
1505 	struct namecache *rncp;
1506 
1507 	cache_assert_bucket_unlocked(ncp);
1508 
1509 	cache_sort_vnodes(&dvlp, &vlp);
1510 	cache_lock_vnodes(dvlp, vlp);
1511 	mtx_lock(blp);
1512 	CK_SLIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1513 		if (rncp == ncp && rncp->nc_dvp == dvp &&
1514 		    rncp->nc_nlen == cnp->cn_namelen &&
1515 		    !bcmp(rncp->nc_name, cnp->cn_nameptr, rncp->nc_nlen))
1516 			break;
1517 	}
1518 	if (rncp != NULL) {
1519 		cache_zap_locked(rncp);
1520 		mtx_unlock(blp);
1521 		cache_unlock_vnodes(dvlp, vlp);
1522 		counter_u64_add(zap_bucket_relock_success, 1);
1523 		return (0);
1524 	}
1525 
1526 	mtx_unlock(blp);
1527 	cache_unlock_vnodes(dvlp, vlp);
1528 	return (EAGAIN);
1529 }
1530 
1531 static int __noinline
1532 cache_zap_locked_bucket(struct namecache *ncp, struct componentname *cnp,
1533     uint32_t hash, struct mtx *blp)
1534 {
1535 	struct mtx *dvlp, *vlp;
1536 	struct vnode *dvp;
1537 
1538 	cache_assert_bucket_locked(ncp);
1539 
1540 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1541 	vlp = NULL;
1542 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1543 		vlp = VP2VNODELOCK(ncp->nc_vp);
1544 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1545 		cache_zap_locked(ncp);
1546 		mtx_unlock(blp);
1547 		cache_unlock_vnodes(dvlp, vlp);
1548 		return (0);
1549 	}
1550 
1551 	dvp = ncp->nc_dvp;
1552 	mtx_unlock(blp);
1553 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1554 }
1555 
1556 static __noinline int
1557 cache_remove_cnp(struct vnode *dvp, struct componentname *cnp)
1558 {
1559 	struct namecache *ncp;
1560 	struct mtx *blp;
1561 	struct mtx *dvlp, *dvlp2;
1562 	uint32_t hash;
1563 	int error;
1564 
1565 	if (cnp->cn_namelen == 2 &&
1566 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1567 		dvlp = VP2VNODELOCK(dvp);
1568 		dvlp2 = NULL;
1569 		mtx_lock(dvlp);
1570 retry_dotdot:
1571 		ncp = dvp->v_cache_dd;
1572 		if (ncp == NULL) {
1573 			mtx_unlock(dvlp);
1574 			if (dvlp2 != NULL)
1575 				mtx_unlock(dvlp2);
1576 			SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1577 			return (0);
1578 		}
1579 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1580 			if (!cache_zap_locked_vnode_kl2(ncp, dvp, &dvlp2))
1581 				goto retry_dotdot;
1582 			MPASS(dvp->v_cache_dd == NULL);
1583 			mtx_unlock(dvlp);
1584 			if (dvlp2 != NULL)
1585 				mtx_unlock(dvlp2);
1586 			cache_free(ncp);
1587 		} else {
1588 			vn_seqc_write_begin(dvp);
1589 			dvp->v_cache_dd = NULL;
1590 			vn_seqc_write_end(dvp);
1591 			mtx_unlock(dvlp);
1592 			if (dvlp2 != NULL)
1593 				mtx_unlock(dvlp2);
1594 		}
1595 		SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1596 		return (1);
1597 	}
1598 
1599 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1600 	blp = HASH2BUCKETLOCK(hash);
1601 retry:
1602 	if (CK_SLIST_EMPTY(NCHHASH(hash)))
1603 		goto out_no_entry;
1604 
1605 	mtx_lock(blp);
1606 
1607 	CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1608 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1609 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1610 			break;
1611 	}
1612 
1613 	if (ncp == NULL) {
1614 		mtx_unlock(blp);
1615 		goto out_no_entry;
1616 	}
1617 
1618 	error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
1619 	if (__predict_false(error != 0)) {
1620 		zap_bucket_fail++;
1621 		goto retry;
1622 	}
1623 	counter_u64_add(numposzaps, 1);
1624 	SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1625 	cache_free(ncp);
1626 	return (1);
1627 out_no_entry:
1628 	counter_u64_add(nummisszap, 1);
1629 	SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1630 	return (0);
1631 }
1632 
1633 static int __noinline
1634 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1635     struct timespec *tsp, int *ticksp)
1636 {
1637 	int ltype;
1638 
1639 	*vpp = dvp;
1640 	counter_u64_add(dothits, 1);
1641 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1642 	if (tsp != NULL)
1643 		timespecclear(tsp);
1644 	if (ticksp != NULL)
1645 		*ticksp = ticks;
1646 	vrefact(*vpp);
1647 	/*
1648 	 * When we lookup "." we still can be asked to lock it
1649 	 * differently...
1650 	 */
1651 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1652 	if (ltype != VOP_ISLOCKED(*vpp)) {
1653 		if (ltype == LK_EXCLUSIVE) {
1654 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1655 			if (VN_IS_DOOMED((*vpp))) {
1656 				/* forced unmount */
1657 				vrele(*vpp);
1658 				*vpp = NULL;
1659 				return (ENOENT);
1660 			}
1661 		} else
1662 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1663 	}
1664 	return (-1);
1665 }
1666 
1667 static int __noinline
1668 cache_lookup_dotdot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1669     struct timespec *tsp, int *ticksp)
1670 {
1671 	struct namecache_ts *ncp_ts;
1672 	struct namecache *ncp;
1673 	struct mtx *dvlp;
1674 	enum vgetstate vs;
1675 	int error, ltype;
1676 	bool whiteout;
1677 
1678 	MPASS((cnp->cn_flags & ISDOTDOT) != 0);
1679 
1680 	if ((cnp->cn_flags & MAKEENTRY) == 0) {
1681 		cache_remove_cnp(dvp, cnp);
1682 		return (0);
1683 	}
1684 
1685 	counter_u64_add(dotdothits, 1);
1686 retry:
1687 	dvlp = VP2VNODELOCK(dvp);
1688 	mtx_lock(dvlp);
1689 	ncp = dvp->v_cache_dd;
1690 	if (ncp == NULL) {
1691 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, "..", NULL);
1692 		mtx_unlock(dvlp);
1693 		return (0);
1694 	}
1695 	if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1696 		if (ncp->nc_flag & NCF_NEGATIVE)
1697 			*vpp = NULL;
1698 		else
1699 			*vpp = ncp->nc_vp;
1700 	} else
1701 		*vpp = ncp->nc_dvp;
1702 	if (*vpp == NULL)
1703 		goto negative_success;
1704 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..", *vpp);
1705 	cache_out_ts(ncp, tsp, ticksp);
1706 	if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
1707 	    NCF_DTS && tsp != NULL) {
1708 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
1709 		*tsp = ncp_ts->nc_dotdottime;
1710 	}
1711 
1712 	MPASS(dvp != *vpp);
1713 	ltype = VOP_ISLOCKED(dvp);
1714 	VOP_UNLOCK(dvp);
1715 	vs = vget_prep(*vpp);
1716 	mtx_unlock(dvlp);
1717 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1718 	vn_lock(dvp, ltype | LK_RETRY);
1719 	if (VN_IS_DOOMED(dvp)) {
1720 		if (error == 0)
1721 			vput(*vpp);
1722 		*vpp = NULL;
1723 		return (ENOENT);
1724 	}
1725 	if (error) {
1726 		*vpp = NULL;
1727 		goto retry;
1728 	}
1729 	return (-1);
1730 negative_success:
1731 	if (__predict_false(cnp->cn_nameiop == CREATE)) {
1732 		if (cnp->cn_flags & ISLASTCN) {
1733 			counter_u64_add(numnegzaps, 1);
1734 			cache_zap_negative_locked_vnode_kl(ncp, dvp);
1735 			mtx_unlock(dvlp);
1736 			cache_free(ncp);
1737 			return (0);
1738 		}
1739 	}
1740 
1741 	whiteout = (ncp->nc_flag & NCF_WHITE);
1742 	cache_out_ts(ncp, tsp, ticksp);
1743 	if (cache_neg_hit_prep(ncp))
1744 		cache_neg_promote(ncp);
1745 	else
1746 		cache_neg_hit_finish(ncp);
1747 	mtx_unlock(dvlp);
1748 	if (whiteout)
1749 		cnp->cn_flags |= ISWHITEOUT;
1750 	return (ENOENT);
1751 }
1752 
1753 /**
1754  * Lookup a name in the name cache
1755  *
1756  * # Arguments
1757  *
1758  * - dvp:	Parent directory in which to search.
1759  * - vpp:	Return argument.  Will contain desired vnode on cache hit.
1760  * - cnp:	Parameters of the name search.  The most interesting bits of
1761  *   		the cn_flags field have the following meanings:
1762  *   	- MAKEENTRY:	If clear, free an entry from the cache rather than look
1763  *   			it up.
1764  *   	- ISDOTDOT:	Must be set if and only if cn_nameptr == ".."
1765  * - tsp:	Return storage for cache timestamp.  On a successful (positive
1766  *   		or negative) lookup, tsp will be filled with any timespec that
1767  *   		was stored when this cache entry was created.  However, it will
1768  *   		be clear for "." entries.
1769  * - ticks:	Return storage for alternate cache timestamp.  On a successful
1770  *   		(positive or negative) lookup, it will contain the ticks value
1771  *   		that was current when the cache entry was created, unless cnp
1772  *   		was ".".
1773  *
1774  * Either both tsp and ticks have to be provided or neither of them.
1775  *
1776  * # Returns
1777  *
1778  * - -1:	A positive cache hit.  vpp will contain the desired vnode.
1779  * - ENOENT:	A negative cache hit, or dvp was recycled out from under us due
1780  *		to a forced unmount.  vpp will not be modified.  If the entry
1781  *		is a whiteout, then the ISWHITEOUT flag will be set in
1782  *		cnp->cn_flags.
1783  * - 0:		A cache miss.  vpp will not be modified.
1784  *
1785  * # Locking
1786  *
1787  * On a cache hit, vpp will be returned locked and ref'd.  If we're looking up
1788  * .., dvp is unlocked.  If we're looking up . an extra ref is taken, but the
1789  * lock is not recursively acquired.
1790  */
1791 static int __noinline
1792 cache_lookup_fallback(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1793     struct timespec *tsp, int *ticksp)
1794 {
1795 	struct namecache *ncp;
1796 	struct mtx *blp;
1797 	uint32_t hash;
1798 	enum vgetstate vs;
1799 	int error;
1800 	bool whiteout;
1801 
1802 	MPASS((cnp->cn_flags & ISDOTDOT) == 0);
1803 	MPASS((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) != 0);
1804 
1805 retry:
1806 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1807 	blp = HASH2BUCKETLOCK(hash);
1808 	mtx_lock(blp);
1809 
1810 	CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1811 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1812 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1813 			break;
1814 	}
1815 
1816 	if (__predict_false(ncp == NULL)) {
1817 		mtx_unlock(blp);
1818 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1819 		    NULL);
1820 		counter_u64_add(nummiss, 1);
1821 		return (0);
1822 	}
1823 
1824 	if (ncp->nc_flag & NCF_NEGATIVE)
1825 		goto negative_success;
1826 
1827 	counter_u64_add(numposhits, 1);
1828 	*vpp = ncp->nc_vp;
1829 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
1830 	cache_out_ts(ncp, tsp, ticksp);
1831 	MPASS(dvp != *vpp);
1832 	vs = vget_prep(*vpp);
1833 	mtx_unlock(blp);
1834 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1835 	if (error) {
1836 		*vpp = NULL;
1837 		goto retry;
1838 	}
1839 	return (-1);
1840 negative_success:
1841 	/*
1842 	 * We don't get here with regular lookup apart from corner cases.
1843 	 */
1844 	if (__predict_true(cnp->cn_nameiop == CREATE)) {
1845 		if (cnp->cn_flags & ISLASTCN) {
1846 			counter_u64_add(numnegzaps, 1);
1847 			error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
1848 			if (__predict_false(error != 0)) {
1849 				zap_bucket_fail2++;
1850 				goto retry;
1851 			}
1852 			cache_free(ncp);
1853 			return (0);
1854 		}
1855 	}
1856 
1857 	whiteout = (ncp->nc_flag & NCF_WHITE);
1858 	cache_out_ts(ncp, tsp, ticksp);
1859 	if (cache_neg_hit_prep(ncp))
1860 		cache_neg_promote(ncp);
1861 	else
1862 		cache_neg_hit_finish(ncp);
1863 	mtx_unlock(blp);
1864 	if (whiteout)
1865 		cnp->cn_flags |= ISWHITEOUT;
1866 	return (ENOENT);
1867 }
1868 
1869 int
1870 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1871     struct timespec *tsp, int *ticksp)
1872 {
1873 	struct namecache *ncp;
1874 	uint32_t hash;
1875 	enum vgetstate vs;
1876 	int error;
1877 	bool whiteout, neg_promote;
1878 	u_short nc_flag;
1879 
1880 	MPASS((tsp == NULL && ticksp == NULL) || (tsp != NULL && ticksp != NULL));
1881 
1882 #ifdef DEBUG_CACHE
1883 	if (__predict_false(!doingcache)) {
1884 		cnp->cn_flags &= ~MAKEENTRY;
1885 		return (0);
1886 	}
1887 #endif
1888 
1889 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
1890 		if (cnp->cn_namelen == 1)
1891 			return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
1892 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.')
1893 			return (cache_lookup_dotdot(dvp, vpp, cnp, tsp, ticksp));
1894 	}
1895 
1896 	MPASS((cnp->cn_flags & ISDOTDOT) == 0);
1897 
1898 	if ((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) == 0) {
1899 		cache_remove_cnp(dvp, cnp);
1900 		return (0);
1901 	}
1902 
1903 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1904 	vfs_smr_enter();
1905 
1906 	CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1907 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
1908 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
1909 			break;
1910 	}
1911 
1912 	if (__predict_false(ncp == NULL)) {
1913 		vfs_smr_exit();
1914 		SDT_PROBE3(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr,
1915 		    NULL);
1916 		counter_u64_add(nummiss, 1);
1917 		return (0);
1918 	}
1919 
1920 	nc_flag = atomic_load_char(&ncp->nc_flag);
1921 	if (nc_flag & NCF_NEGATIVE)
1922 		goto negative_success;
1923 
1924 	counter_u64_add(numposhits, 1);
1925 	*vpp = ncp->nc_vp;
1926 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
1927 	cache_out_ts(ncp, tsp, ticksp);
1928 	MPASS(dvp != *vpp);
1929 	if (!cache_ncp_canuse(ncp)) {
1930 		vfs_smr_exit();
1931 		*vpp = NULL;
1932 		goto out_fallback;
1933 	}
1934 	vs = vget_prep_smr(*vpp);
1935 	vfs_smr_exit();
1936 	if (__predict_false(vs == VGET_NONE)) {
1937 		*vpp = NULL;
1938 		goto out_fallback;
1939 	}
1940 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
1941 	if (error) {
1942 		*vpp = NULL;
1943 		goto out_fallback;
1944 	}
1945 	return (-1);
1946 negative_success:
1947 	if (cnp->cn_nameiop == CREATE) {
1948 		if (cnp->cn_flags & ISLASTCN) {
1949 			vfs_smr_exit();
1950 			goto out_fallback;
1951 		}
1952 	}
1953 
1954 	cache_out_ts(ncp, tsp, ticksp);
1955 	whiteout = (ncp->nc_flag & NCF_WHITE);
1956 	neg_promote = cache_neg_hit_prep(ncp);
1957 	if (!cache_ncp_canuse(ncp)) {
1958 		cache_neg_hit_abort(ncp);
1959 		vfs_smr_exit();
1960 		goto out_fallback;
1961 	}
1962 	if (neg_promote) {
1963 		vfs_smr_exit();
1964 		if (!cache_neg_promote_cond(dvp, cnp, ncp, hash))
1965 			goto out_fallback;
1966 	} else {
1967 		cache_neg_hit_finish(ncp);
1968 		vfs_smr_exit();
1969 	}
1970 	if (whiteout)
1971 		cnp->cn_flags |= ISWHITEOUT;
1972 	return (ENOENT);
1973 out_fallback:
1974 	return (cache_lookup_fallback(dvp, vpp, cnp, tsp, ticksp));
1975 }
1976 
1977 struct celockstate {
1978 	struct mtx *vlp[3];
1979 	struct mtx *blp[2];
1980 };
1981 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
1982 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
1983 
1984 static inline void
1985 cache_celockstate_init(struct celockstate *cel)
1986 {
1987 
1988 	bzero(cel, sizeof(*cel));
1989 }
1990 
1991 static void
1992 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
1993     struct vnode *dvp)
1994 {
1995 	struct mtx *vlp1, *vlp2;
1996 
1997 	MPASS(cel->vlp[0] == NULL);
1998 	MPASS(cel->vlp[1] == NULL);
1999 	MPASS(cel->vlp[2] == NULL);
2000 
2001 	MPASS(vp != NULL || dvp != NULL);
2002 
2003 	vlp1 = VP2VNODELOCK(vp);
2004 	vlp2 = VP2VNODELOCK(dvp);
2005 	cache_sort_vnodes(&vlp1, &vlp2);
2006 
2007 	if (vlp1 != NULL) {
2008 		mtx_lock(vlp1);
2009 		cel->vlp[0] = vlp1;
2010 	}
2011 	mtx_lock(vlp2);
2012 	cel->vlp[1] = vlp2;
2013 }
2014 
2015 static void
2016 cache_unlock_vnodes_cel(struct celockstate *cel)
2017 {
2018 
2019 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
2020 
2021 	if (cel->vlp[0] != NULL)
2022 		mtx_unlock(cel->vlp[0]);
2023 	if (cel->vlp[1] != NULL)
2024 		mtx_unlock(cel->vlp[1]);
2025 	if (cel->vlp[2] != NULL)
2026 		mtx_unlock(cel->vlp[2]);
2027 }
2028 
2029 static bool
2030 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
2031 {
2032 	struct mtx *vlp;
2033 	bool ret;
2034 
2035 	cache_assert_vlp_locked(cel->vlp[0]);
2036 	cache_assert_vlp_locked(cel->vlp[1]);
2037 	MPASS(cel->vlp[2] == NULL);
2038 
2039 	MPASS(vp != NULL);
2040 	vlp = VP2VNODELOCK(vp);
2041 
2042 	ret = true;
2043 	if (vlp >= cel->vlp[1]) {
2044 		mtx_lock(vlp);
2045 	} else {
2046 		if (mtx_trylock(vlp))
2047 			goto out;
2048 		cache_lock_vnodes_cel_3_failures++;
2049 		cache_unlock_vnodes_cel(cel);
2050 		if (vlp < cel->vlp[0]) {
2051 			mtx_lock(vlp);
2052 			mtx_lock(cel->vlp[0]);
2053 			mtx_lock(cel->vlp[1]);
2054 		} else {
2055 			if (cel->vlp[0] != NULL)
2056 				mtx_lock(cel->vlp[0]);
2057 			mtx_lock(vlp);
2058 			mtx_lock(cel->vlp[1]);
2059 		}
2060 		ret = false;
2061 	}
2062 out:
2063 	cel->vlp[2] = vlp;
2064 	return (ret);
2065 }
2066 
2067 static void
2068 cache_lock_buckets_cel(struct celockstate *cel, struct mtx *blp1,
2069     struct mtx *blp2)
2070 {
2071 
2072 	MPASS(cel->blp[0] == NULL);
2073 	MPASS(cel->blp[1] == NULL);
2074 
2075 	cache_sort_vnodes(&blp1, &blp2);
2076 
2077 	if (blp1 != NULL) {
2078 		mtx_lock(blp1);
2079 		cel->blp[0] = blp1;
2080 	}
2081 	mtx_lock(blp2);
2082 	cel->blp[1] = blp2;
2083 }
2084 
2085 static void
2086 cache_unlock_buckets_cel(struct celockstate *cel)
2087 {
2088 
2089 	if (cel->blp[0] != NULL)
2090 		mtx_unlock(cel->blp[0]);
2091 	mtx_unlock(cel->blp[1]);
2092 }
2093 
2094 /*
2095  * Lock part of the cache affected by the insertion.
2096  *
2097  * This means vnodelocks for dvp, vp and the relevant bucketlock.
2098  * However, insertion can result in removal of an old entry. In this
2099  * case we have an additional vnode and bucketlock pair to lock.
2100  *
2101  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
2102  * preserving the locking order (smaller address first).
2103  */
2104 static void
2105 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2106     uint32_t hash)
2107 {
2108 	struct namecache *ncp;
2109 	struct mtx *blps[2];
2110 
2111 	blps[0] = HASH2BUCKETLOCK(hash);
2112 	for (;;) {
2113 		blps[1] = NULL;
2114 		cache_lock_vnodes_cel(cel, dvp, vp);
2115 		if (vp == NULL || vp->v_type != VDIR)
2116 			break;
2117 		ncp = vp->v_cache_dd;
2118 		if (ncp == NULL)
2119 			break;
2120 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2121 			break;
2122 		MPASS(ncp->nc_dvp == vp);
2123 		blps[1] = NCP2BUCKETLOCK(ncp);
2124 		if (ncp->nc_flag & NCF_NEGATIVE)
2125 			break;
2126 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2127 			break;
2128 		/*
2129 		 * All vnodes got re-locked. Re-validate the state and if
2130 		 * nothing changed we are done. Otherwise restart.
2131 		 */
2132 		if (ncp == vp->v_cache_dd &&
2133 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2134 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
2135 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2136 			break;
2137 		cache_unlock_vnodes_cel(cel);
2138 		cel->vlp[0] = NULL;
2139 		cel->vlp[1] = NULL;
2140 		cel->vlp[2] = NULL;
2141 	}
2142 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
2143 }
2144 
2145 static void
2146 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2147     uint32_t hash)
2148 {
2149 	struct namecache *ncp;
2150 	struct mtx *blps[2];
2151 
2152 	blps[0] = HASH2BUCKETLOCK(hash);
2153 	for (;;) {
2154 		blps[1] = NULL;
2155 		cache_lock_vnodes_cel(cel, dvp, vp);
2156 		ncp = dvp->v_cache_dd;
2157 		if (ncp == NULL)
2158 			break;
2159 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2160 			break;
2161 		MPASS(ncp->nc_dvp == dvp);
2162 		blps[1] = NCP2BUCKETLOCK(ncp);
2163 		if (ncp->nc_flag & NCF_NEGATIVE)
2164 			break;
2165 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2166 			break;
2167 		if (ncp == dvp->v_cache_dd &&
2168 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2169 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
2170 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2171 			break;
2172 		cache_unlock_vnodes_cel(cel);
2173 		cel->vlp[0] = NULL;
2174 		cel->vlp[1] = NULL;
2175 		cel->vlp[2] = NULL;
2176 	}
2177 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
2178 }
2179 
2180 static void
2181 cache_enter_unlock(struct celockstate *cel)
2182 {
2183 
2184 	cache_unlock_buckets_cel(cel);
2185 	cache_unlock_vnodes_cel(cel);
2186 }
2187 
2188 static void __noinline
2189 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp,
2190     struct componentname *cnp)
2191 {
2192 	struct celockstate cel;
2193 	struct namecache *ncp;
2194 	uint32_t hash;
2195 	int len;
2196 
2197 	if (dvp->v_cache_dd == NULL)
2198 		return;
2199 	len = cnp->cn_namelen;
2200 	cache_celockstate_init(&cel);
2201 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2202 	cache_enter_lock_dd(&cel, dvp, vp, hash);
2203 	vn_seqc_write_begin(dvp);
2204 	ncp = dvp->v_cache_dd;
2205 	if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) {
2206 		KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent"));
2207 		cache_zap_locked(ncp);
2208 	} else {
2209 		ncp = NULL;
2210 	}
2211 	dvp->v_cache_dd = NULL;
2212 	vn_seqc_write_end(dvp);
2213 	cache_enter_unlock(&cel);
2214 	if (ncp != NULL)
2215 		cache_free(ncp);
2216 }
2217 
2218 /*
2219  * Add an entry to the cache.
2220  */
2221 void
2222 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
2223     struct timespec *tsp, struct timespec *dtsp)
2224 {
2225 	struct celockstate cel;
2226 	struct namecache *ncp, *n2, *ndd;
2227 	struct namecache_ts *ncp_ts;
2228 	struct nchashhead *ncpp;
2229 	uint32_t hash;
2230 	int flag;
2231 	int len;
2232 
2233 	KASSERT(cnp->cn_namelen <= NAME_MAX,
2234 	    ("%s: passed len %ld exceeds NAME_MAX (%d)", __func__, cnp->cn_namelen,
2235 	    NAME_MAX));
2236 	VNPASS(dvp != vp, dvp);
2237 	VNPASS(!VN_IS_DOOMED(dvp), dvp);
2238 	VNPASS(dvp->v_type != VNON, dvp);
2239 	if (vp != NULL) {
2240 		VNPASS(!VN_IS_DOOMED(vp), vp);
2241 		VNPASS(vp->v_type != VNON, vp);
2242 	}
2243 
2244 #ifdef DEBUG_CACHE
2245 	if (__predict_false(!doingcache))
2246 		return;
2247 #endif
2248 
2249 	flag = 0;
2250 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
2251 		if (cnp->cn_namelen == 1)
2252 			return;
2253 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
2254 			cache_enter_dotdot_prep(dvp, vp, cnp);
2255 			flag = NCF_ISDOTDOT;
2256 		}
2257 	}
2258 
2259 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
2260 	if (ncp == NULL)
2261 		return;
2262 
2263 	cache_celockstate_init(&cel);
2264 	ndd = NULL;
2265 	ncp_ts = NULL;
2266 
2267 	/*
2268 	 * Calculate the hash key and setup as much of the new
2269 	 * namecache entry as possible before acquiring the lock.
2270 	 */
2271 	ncp->nc_flag = flag | NCF_WIP;
2272 	ncp->nc_vp = vp;
2273 	if (vp == NULL)
2274 		cache_neg_init(ncp);
2275 	ncp->nc_dvp = dvp;
2276 	if (tsp != NULL) {
2277 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
2278 		ncp_ts->nc_time = *tsp;
2279 		ncp_ts->nc_ticks = ticks;
2280 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
2281 		if (dtsp != NULL) {
2282 			ncp_ts->nc_dotdottime = *dtsp;
2283 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
2284 		}
2285 	}
2286 	len = ncp->nc_nlen = cnp->cn_namelen;
2287 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2288 	memcpy(ncp->nc_name, cnp->cn_nameptr, len);
2289 	ncp->nc_name[len] = '\0';
2290 	cache_enter_lock(&cel, dvp, vp, hash);
2291 
2292 	/*
2293 	 * See if this vnode or negative entry is already in the cache
2294 	 * with this name.  This can happen with concurrent lookups of
2295 	 * the same path name.
2296 	 */
2297 	ncpp = NCHHASH(hash);
2298 	CK_SLIST_FOREACH(n2, ncpp, nc_hash) {
2299 		if (n2->nc_dvp == dvp &&
2300 		    n2->nc_nlen == cnp->cn_namelen &&
2301 		    !bcmp(n2->nc_name, cnp->cn_nameptr, n2->nc_nlen)) {
2302 			MPASS(cache_ncp_canuse(n2));
2303 			if ((n2->nc_flag & NCF_NEGATIVE) != 0)
2304 				KASSERT(vp == NULL,
2305 				    ("%s: found entry pointing to a different vnode (%p != %p)",
2306 				    __func__, NULL, vp));
2307 			else
2308 				KASSERT(n2->nc_vp == vp,
2309 				    ("%s: found entry pointing to a different vnode (%p != %p)",
2310 				    __func__, n2->nc_vp, vp));
2311 			/*
2312 			 * Entries are supposed to be immutable unless in the
2313 			 * process of getting destroyed. Accommodating for
2314 			 * changing timestamps is possible but not worth it.
2315 			 * This should be harmless in terms of correctness, in
2316 			 * the worst case resulting in an earlier expiration.
2317 			 * Alternatively, the found entry can be replaced
2318 			 * altogether.
2319 			 */
2320 			MPASS((n2->nc_flag & (NCF_TS | NCF_DTS)) == (ncp->nc_flag & (NCF_TS | NCF_DTS)));
2321 #if 0
2322 			if (tsp != NULL) {
2323 				KASSERT((n2->nc_flag & NCF_TS) != 0,
2324 				    ("no NCF_TS"));
2325 				n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
2326 				n2_ts->nc_time = ncp_ts->nc_time;
2327 				n2_ts->nc_ticks = ncp_ts->nc_ticks;
2328 				if (dtsp != NULL) {
2329 					n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
2330 					n2_ts->nc_nc.nc_flag |= NCF_DTS;
2331 				}
2332 			}
2333 #endif
2334 			SDT_PROBE3(vfs, namecache, enter, duplicate, dvp, ncp->nc_name,
2335 			    vp);
2336 			goto out_unlock_free;
2337 		}
2338 	}
2339 
2340 	if (flag == NCF_ISDOTDOT) {
2341 		/*
2342 		 * See if we are trying to add .. entry, but some other lookup
2343 		 * has populated v_cache_dd pointer already.
2344 		 */
2345 		if (dvp->v_cache_dd != NULL)
2346 			goto out_unlock_free;
2347 		KASSERT(vp == NULL || vp->v_type == VDIR,
2348 		    ("wrong vnode type %p", vp));
2349 		vn_seqc_write_begin(dvp);
2350 		dvp->v_cache_dd = ncp;
2351 		vn_seqc_write_end(dvp);
2352 	}
2353 
2354 	if (vp != NULL) {
2355 		if (flag != NCF_ISDOTDOT) {
2356 			/*
2357 			 * For this case, the cache entry maps both the
2358 			 * directory name in it and the name ".." for the
2359 			 * directory's parent.
2360 			 */
2361 			vn_seqc_write_begin(vp);
2362 			if ((ndd = vp->v_cache_dd) != NULL) {
2363 				if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
2364 					cache_zap_locked(ndd);
2365 				else
2366 					ndd = NULL;
2367 			}
2368 			vp->v_cache_dd = ncp;
2369 			vn_seqc_write_end(vp);
2370 		} else if (vp->v_type != VDIR) {
2371 			if (vp->v_cache_dd != NULL) {
2372 				vn_seqc_write_begin(vp);
2373 				vp->v_cache_dd = NULL;
2374 				vn_seqc_write_end(vp);
2375 			}
2376 		}
2377 	}
2378 
2379 	if (flag != NCF_ISDOTDOT) {
2380 		if (LIST_EMPTY(&dvp->v_cache_src)) {
2381 			cache_hold_vnode(dvp);
2382 		}
2383 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
2384 	}
2385 
2386 	/*
2387 	 * If the entry is "negative", we place it into the
2388 	 * "negative" cache queue, otherwise, we place it into the
2389 	 * destination vnode's cache entries queue.
2390 	 */
2391 	if (vp != NULL) {
2392 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
2393 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
2394 		    vp);
2395 	} else {
2396 		if (cnp->cn_flags & ISWHITEOUT)
2397 			ncp->nc_flag |= NCF_WHITE;
2398 		cache_neg_insert(ncp);
2399 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
2400 		    ncp->nc_name);
2401 	}
2402 
2403 	/*
2404 	 * Insert the new namecache entry into the appropriate chain
2405 	 * within the cache entries table.
2406 	 */
2407 	CK_SLIST_INSERT_HEAD(ncpp, ncp, nc_hash);
2408 
2409 	atomic_thread_fence_rel();
2410 	/*
2411 	 * Mark the entry as fully constructed.
2412 	 * It is immutable past this point until its removal.
2413 	 */
2414 	atomic_store_char(&ncp->nc_flag, ncp->nc_flag & ~NCF_WIP);
2415 
2416 	cache_enter_unlock(&cel);
2417 	if (ndd != NULL)
2418 		cache_free(ndd);
2419 	return;
2420 out_unlock_free:
2421 	cache_enter_unlock(&cel);
2422 	cache_free(ncp);
2423 	return;
2424 }
2425 
2426 static u_int
2427 cache_roundup_2(u_int val)
2428 {
2429 	u_int res;
2430 
2431 	for (res = 1; res <= val; res <<= 1)
2432 		continue;
2433 
2434 	return (res);
2435 }
2436 
2437 static struct nchashhead *
2438 nchinittbl(u_long elements, u_long *hashmask)
2439 {
2440 	struct nchashhead *hashtbl;
2441 	u_long hashsize, i;
2442 
2443 	hashsize = cache_roundup_2(elements) / 2;
2444 
2445 	hashtbl = malloc((u_long)hashsize * sizeof(*hashtbl), M_VFSCACHE, M_WAITOK);
2446 	for (i = 0; i < hashsize; i++)
2447 		CK_SLIST_INIT(&hashtbl[i]);
2448 	*hashmask = hashsize - 1;
2449 	return (hashtbl);
2450 }
2451 
2452 static void
2453 ncfreetbl(struct nchashhead *hashtbl)
2454 {
2455 
2456 	free(hashtbl, M_VFSCACHE);
2457 }
2458 
2459 /*
2460  * Name cache initialization, from vfs_init() when we are booting
2461  */
2462 static void
2463 nchinit(void *dummy __unused)
2464 {
2465 	u_int i;
2466 
2467 	cache_zone_small = uma_zcreate("S VFS Cache", CACHE_ZONE_SMALL_SIZE,
2468 	    NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2469 	cache_zone_small_ts = uma_zcreate("STS VFS Cache", CACHE_ZONE_SMALL_TS_SIZE,
2470 	    NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2471 	cache_zone_large = uma_zcreate("L VFS Cache", CACHE_ZONE_LARGE_SIZE,
2472 	    NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2473 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache", CACHE_ZONE_LARGE_TS_SIZE,
2474 	    NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2475 
2476 	VFS_SMR_ZONE_SET(cache_zone_small);
2477 	VFS_SMR_ZONE_SET(cache_zone_small_ts);
2478 	VFS_SMR_ZONE_SET(cache_zone_large);
2479 	VFS_SMR_ZONE_SET(cache_zone_large_ts);
2480 
2481 	ncsize = desiredvnodes * ncsizefactor;
2482 	cache_recalc_neg_min(ncnegminpct);
2483 	nchashtbl = nchinittbl(desiredvnodes * 2, &nchash);
2484 	ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
2485 	if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
2486 		ncbuckethash = 7;
2487 	if (ncbuckethash > nchash)
2488 		ncbuckethash = nchash;
2489 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
2490 	    M_WAITOK | M_ZERO);
2491 	for (i = 0; i < numbucketlocks; i++)
2492 		mtx_init(&bucketlocks[i], "ncbuc", NULL, MTX_DUPOK | MTX_RECURSE);
2493 	ncvnodehash = ncbuckethash;
2494 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
2495 	    M_WAITOK | M_ZERO);
2496 	for (i = 0; i < numvnodelocks; i++)
2497 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
2498 
2499 	for (i = 0; i < numneglists; i++) {
2500 		mtx_init(&neglists[i].nl_evict_lock, "ncnege", NULL, MTX_DEF);
2501 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
2502 		TAILQ_INIT(&neglists[i].nl_list);
2503 		TAILQ_INIT(&neglists[i].nl_hotlist);
2504 	}
2505 }
2506 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
2507 
2508 void
2509 cache_vnode_init(struct vnode *vp)
2510 {
2511 
2512 	LIST_INIT(&vp->v_cache_src);
2513 	TAILQ_INIT(&vp->v_cache_dst);
2514 	vp->v_cache_dd = NULL;
2515 	cache_prehash(vp);
2516 }
2517 
2518 void
2519 cache_changesize(u_long newmaxvnodes)
2520 {
2521 	struct nchashhead *new_nchashtbl, *old_nchashtbl;
2522 	u_long new_nchash, old_nchash;
2523 	struct namecache *ncp;
2524 	uint32_t hash;
2525 	u_long newncsize;
2526 	int i;
2527 
2528 	newncsize = newmaxvnodes * ncsizefactor;
2529 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
2530 	if (newmaxvnodes < numbucketlocks)
2531 		newmaxvnodes = numbucketlocks;
2532 
2533 	new_nchashtbl = nchinittbl(newmaxvnodes, &new_nchash);
2534 	/* If same hash table size, nothing to do */
2535 	if (nchash == new_nchash) {
2536 		ncfreetbl(new_nchashtbl);
2537 		return;
2538 	}
2539 	/*
2540 	 * Move everything from the old hash table to the new table.
2541 	 * None of the namecache entries in the table can be removed
2542 	 * because to do so, they have to be removed from the hash table.
2543 	 */
2544 	cache_lock_all_vnodes();
2545 	cache_lock_all_buckets();
2546 	old_nchashtbl = nchashtbl;
2547 	old_nchash = nchash;
2548 	nchashtbl = new_nchashtbl;
2549 	nchash = new_nchash;
2550 	for (i = 0; i <= old_nchash; i++) {
2551 		while ((ncp = CK_SLIST_FIRST(&old_nchashtbl[i])) != NULL) {
2552 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
2553 			    ncp->nc_dvp);
2554 			CK_SLIST_REMOVE(&old_nchashtbl[i], ncp, namecache, nc_hash);
2555 			CK_SLIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
2556 		}
2557 	}
2558 	ncsize = newncsize;
2559 	cache_recalc_neg_min(ncnegminpct);
2560 	cache_unlock_all_buckets();
2561 	cache_unlock_all_vnodes();
2562 	ncfreetbl(old_nchashtbl);
2563 }
2564 
2565 /*
2566  * Invalidate all entries from and to a particular vnode.
2567  */
2568 static void
2569 cache_purge_impl(struct vnode *vp)
2570 {
2571 	struct cache_freebatch batch;
2572 	struct namecache *ncp;
2573 	struct mtx *vlp, *vlp2;
2574 
2575 	TAILQ_INIT(&batch);
2576 	vlp = VP2VNODELOCK(vp);
2577 	vlp2 = NULL;
2578 	mtx_lock(vlp);
2579 retry:
2580 	while (!LIST_EMPTY(&vp->v_cache_src)) {
2581 		ncp = LIST_FIRST(&vp->v_cache_src);
2582 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2583 			goto retry;
2584 		TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2585 	}
2586 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2587 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
2588 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2589 			goto retry;
2590 		TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2591 	}
2592 	ncp = vp->v_cache_dd;
2593 	if (ncp != NULL) {
2594 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2595 		   ("lost dotdot link"));
2596 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2597 			goto retry;
2598 		TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2599 	}
2600 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2601 	mtx_unlock(vlp);
2602 	if (vlp2 != NULL)
2603 		mtx_unlock(vlp2);
2604 	cache_free_batch(&batch);
2605 }
2606 
2607 /*
2608  * Opportunistic check to see if there is anything to do.
2609  */
2610 static bool
2611 cache_has_entries(struct vnode *vp)
2612 {
2613 
2614 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
2615 	    vp->v_cache_dd == NULL)
2616 		return (false);
2617 	return (true);
2618 }
2619 
2620 void
2621 cache_purge(struct vnode *vp)
2622 {
2623 
2624 	SDT_PROBE1(vfs, namecache, purge, done, vp);
2625 	if (!cache_has_entries(vp))
2626 		return;
2627 	cache_purge_impl(vp);
2628 }
2629 
2630 /*
2631  * Only to be used by vgone.
2632  */
2633 void
2634 cache_purge_vgone(struct vnode *vp)
2635 {
2636 	struct mtx *vlp;
2637 
2638 	VNPASS(VN_IS_DOOMED(vp), vp);
2639 	if (cache_has_entries(vp)) {
2640 		cache_purge_impl(vp);
2641 		return;
2642 	}
2643 
2644 	/*
2645 	 * Serialize against a potential thread doing cache_purge.
2646 	 */
2647 	vlp = VP2VNODELOCK(vp);
2648 	mtx_wait_unlocked(vlp);
2649 	if (cache_has_entries(vp)) {
2650 		cache_purge_impl(vp);
2651 		return;
2652 	}
2653 	return;
2654 }
2655 
2656 /*
2657  * Invalidate all negative entries for a particular directory vnode.
2658  */
2659 void
2660 cache_purge_negative(struct vnode *vp)
2661 {
2662 	struct cache_freebatch batch;
2663 	struct namecache *ncp, *nnp;
2664 	struct mtx *vlp;
2665 
2666 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
2667 	if (LIST_EMPTY(&vp->v_cache_src))
2668 		return;
2669 	TAILQ_INIT(&batch);
2670 	vlp = VP2VNODELOCK(vp);
2671 	mtx_lock(vlp);
2672 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
2673 		if (!(ncp->nc_flag & NCF_NEGATIVE))
2674 			continue;
2675 		cache_zap_negative_locked_vnode_kl(ncp, vp);
2676 		TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2677 	}
2678 	mtx_unlock(vlp);
2679 	cache_free_batch(&batch);
2680 }
2681 
2682 /*
2683  * Entry points for modifying VOP operations.
2684  */
2685 void
2686 cache_vop_rename(struct vnode *fdvp, struct vnode *fvp, struct vnode *tdvp,
2687     struct vnode *tvp, struct componentname *fcnp, struct componentname *tcnp)
2688 {
2689 
2690 	ASSERT_VOP_IN_SEQC(fdvp);
2691 	ASSERT_VOP_IN_SEQC(fvp);
2692 	ASSERT_VOP_IN_SEQC(tdvp);
2693 	if (tvp != NULL)
2694 		ASSERT_VOP_IN_SEQC(tvp);
2695 
2696 	cache_purge(fvp);
2697 	if (tvp != NULL) {
2698 		cache_purge(tvp);
2699 		KASSERT(!cache_remove_cnp(tdvp, tcnp),
2700 		    ("%s: lingering negative entry", __func__));
2701 	} else {
2702 		cache_remove_cnp(tdvp, tcnp);
2703 	}
2704 }
2705 
2706 void
2707 cache_vop_rmdir(struct vnode *dvp, struct vnode *vp)
2708 {
2709 
2710 	ASSERT_VOP_IN_SEQC(dvp);
2711 	ASSERT_VOP_IN_SEQC(vp);
2712 	cache_purge(vp);
2713 }
2714 
2715 #ifdef INVARIANTS
2716 /*
2717  * Validate that if an entry exists it matches.
2718  */
2719 void
2720 cache_validate(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
2721 {
2722 	struct namecache *ncp;
2723 	struct mtx *blp;
2724 	uint32_t hash;
2725 
2726 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
2727 	if (CK_SLIST_EMPTY(NCHHASH(hash)))
2728 		return;
2729 	blp = HASH2BUCKETLOCK(hash);
2730 	mtx_lock(blp);
2731 	CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
2732 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
2733 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen)) {
2734 			if (ncp->nc_vp != vp)
2735 				panic("%s: mismatch (%p != %p); ncp %p [%s] dvp %p vp %p\n",
2736 				    __func__, vp, ncp->nc_vp, ncp, ncp->nc_name, ncp->nc_dvp,
2737 				    ncp->nc_vp);
2738 		}
2739 	}
2740 	mtx_unlock(blp);
2741 }
2742 #endif
2743 
2744 /*
2745  * Flush all entries referencing a particular filesystem.
2746  */
2747 void
2748 cache_purgevfs(struct mount *mp)
2749 {
2750 	struct vnode *vp, *mvp;
2751 
2752 	SDT_PROBE1(vfs, namecache, purgevfs, done, mp);
2753 	/*
2754 	 * Somewhat wasteful iteration over all vnodes. Would be better to
2755 	 * support filtering and avoid the interlock to begin with.
2756 	 */
2757 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2758 		if (!cache_has_entries(vp)) {
2759 			VI_UNLOCK(vp);
2760 			continue;
2761 		}
2762 		vholdl(vp);
2763 		VI_UNLOCK(vp);
2764 		cache_purge(vp);
2765 		vdrop(vp);
2766 	}
2767 }
2768 
2769 /*
2770  * Perform canonical checks and cache lookup and pass on to filesystem
2771  * through the vop_cachedlookup only if needed.
2772  */
2773 
2774 int
2775 vfs_cache_lookup(struct vop_lookup_args *ap)
2776 {
2777 	struct vnode *dvp;
2778 	int error;
2779 	struct vnode **vpp = ap->a_vpp;
2780 	struct componentname *cnp = ap->a_cnp;
2781 	int flags = cnp->cn_flags;
2782 
2783 	*vpp = NULL;
2784 	dvp = ap->a_dvp;
2785 
2786 	if (dvp->v_type != VDIR)
2787 		return (ENOTDIR);
2788 
2789 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
2790 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
2791 		return (EROFS);
2792 
2793 	error = vn_dir_check_exec(dvp, cnp);
2794 	if (error != 0)
2795 		return (error);
2796 
2797 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
2798 	if (error == 0)
2799 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
2800 	if (error == -1)
2801 		return (0);
2802 	return (error);
2803 }
2804 
2805 /* Implementation of the getcwd syscall. */
2806 int
2807 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
2808 {
2809 	char *buf, *retbuf;
2810 	size_t buflen;
2811 	int error;
2812 
2813 	buflen = uap->buflen;
2814 	if (__predict_false(buflen < 2))
2815 		return (EINVAL);
2816 	if (buflen > MAXPATHLEN)
2817 		buflen = MAXPATHLEN;
2818 
2819 	buf = uma_zalloc(namei_zone, M_WAITOK);
2820 	error = vn_getcwd(buf, &retbuf, &buflen);
2821 	if (error == 0)
2822 		error = copyout(retbuf, uap->buf, buflen);
2823 	uma_zfree(namei_zone, buf);
2824 	return (error);
2825 }
2826 
2827 int
2828 vn_getcwd(char *buf, char **retbuf, size_t *buflen)
2829 {
2830 	struct pwd *pwd;
2831 	int error;
2832 
2833 	vfs_smr_enter();
2834 	pwd = pwd_get_smr();
2835 	error = vn_fullpath_any_smr(pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf,
2836 	    buflen, 0);
2837 	VFS_SMR_ASSERT_NOT_ENTERED();
2838 	if (error < 0) {
2839 		pwd = pwd_hold(curthread);
2840 		error = vn_fullpath_any(pwd->pwd_cdir, pwd->pwd_rdir, buf,
2841 		    retbuf, buflen);
2842 		pwd_drop(pwd);
2843 	}
2844 
2845 #ifdef KTRACE
2846 	if (KTRPOINT(curthread, KTR_NAMEI) && error == 0)
2847 		ktrnamei(*retbuf);
2848 #endif
2849 	return (error);
2850 }
2851 
2852 static int
2853 kern___realpathat(struct thread *td, int fd, const char *path, char *buf,
2854     size_t size, int flags, enum uio_seg pathseg)
2855 {
2856 	struct nameidata nd;
2857 	char *retbuf, *freebuf;
2858 	int error;
2859 
2860 	if (flags != 0)
2861 		return (EINVAL);
2862 	NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | SAVENAME | WANTPARENT | AUDITVNODE1,
2863 	    pathseg, path, fd, &cap_fstat_rights, td);
2864 	if ((error = namei(&nd)) != 0)
2865 		return (error);
2866 	error = vn_fullpath_hardlink(&nd, &retbuf, &freebuf, &size);
2867 	if (error == 0) {
2868 		error = copyout(retbuf, buf, size);
2869 		free(freebuf, M_TEMP);
2870 	}
2871 	NDFREE(&nd, 0);
2872 	return (error);
2873 }
2874 
2875 int
2876 sys___realpathat(struct thread *td, struct __realpathat_args *uap)
2877 {
2878 
2879 	return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size,
2880 	    uap->flags, UIO_USERSPACE));
2881 }
2882 
2883 /*
2884  * Retrieve the full filesystem path that correspond to a vnode from the name
2885  * cache (if available)
2886  */
2887 int
2888 vn_fullpath(struct vnode *vp, char **retbuf, char **freebuf)
2889 {
2890 	struct pwd *pwd;
2891 	char *buf;
2892 	size_t buflen;
2893 	int error;
2894 
2895 	if (__predict_false(vp == NULL))
2896 		return (EINVAL);
2897 
2898 	buflen = MAXPATHLEN;
2899 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2900 	vfs_smr_enter();
2901 	pwd = pwd_get_smr();
2902 	error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, &buflen, 0);
2903 	VFS_SMR_ASSERT_NOT_ENTERED();
2904 	if (error < 0) {
2905 		pwd = pwd_hold(curthread);
2906 		error = vn_fullpath_any(vp, pwd->pwd_rdir, buf, retbuf, &buflen);
2907 		pwd_drop(pwd);
2908 	}
2909 	if (error == 0)
2910 		*freebuf = buf;
2911 	else
2912 		free(buf, M_TEMP);
2913 	return (error);
2914 }
2915 
2916 /*
2917  * This function is similar to vn_fullpath, but it attempts to lookup the
2918  * pathname relative to the global root mount point.  This is required for the
2919  * auditing sub-system, as audited pathnames must be absolute, relative to the
2920  * global root mount point.
2921  */
2922 int
2923 vn_fullpath_global(struct vnode *vp, char **retbuf, char **freebuf)
2924 {
2925 	char *buf;
2926 	size_t buflen;
2927 	int error;
2928 
2929 	if (__predict_false(vp == NULL))
2930 		return (EINVAL);
2931 	buflen = MAXPATHLEN;
2932 	buf = malloc(buflen, M_TEMP, M_WAITOK);
2933 	vfs_smr_enter();
2934 	error = vn_fullpath_any_smr(vp, rootvnode, buf, retbuf, &buflen, 0);
2935 	VFS_SMR_ASSERT_NOT_ENTERED();
2936 	if (error < 0) {
2937 		error = vn_fullpath_any(vp, rootvnode, buf, retbuf, &buflen);
2938 	}
2939 	if (error == 0)
2940 		*freebuf = buf;
2941 	else
2942 		free(buf, M_TEMP);
2943 	return (error);
2944 }
2945 
2946 static struct namecache *
2947 vn_dd_from_dst(struct vnode *vp)
2948 {
2949 	struct namecache *ncp;
2950 
2951 	cache_assert_vnode_locked(vp);
2952 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) {
2953 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
2954 			return (ncp);
2955 	}
2956 	return (NULL);
2957 }
2958 
2959 int
2960 vn_vptocnp(struct vnode **vp, char *buf, size_t *buflen)
2961 {
2962 	struct vnode *dvp;
2963 	struct namecache *ncp;
2964 	struct mtx *vlp;
2965 	int error;
2966 
2967 	vlp = VP2VNODELOCK(*vp);
2968 	mtx_lock(vlp);
2969 	ncp = (*vp)->v_cache_dd;
2970 	if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT) == 0) {
2971 		KASSERT(ncp == vn_dd_from_dst(*vp),
2972 		    ("%s: mismatch for dd entry (%p != %p)", __func__,
2973 		    ncp, vn_dd_from_dst(*vp)));
2974 	} else {
2975 		ncp = vn_dd_from_dst(*vp);
2976 	}
2977 	if (ncp != NULL) {
2978 		if (*buflen < ncp->nc_nlen) {
2979 			mtx_unlock(vlp);
2980 			vrele(*vp);
2981 			counter_u64_add(numfullpathfail4, 1);
2982 			error = ENOMEM;
2983 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
2984 			    vp, NULL);
2985 			return (error);
2986 		}
2987 		*buflen -= ncp->nc_nlen;
2988 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
2989 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
2990 		    ncp->nc_name, vp);
2991 		dvp = *vp;
2992 		*vp = ncp->nc_dvp;
2993 		vref(*vp);
2994 		mtx_unlock(vlp);
2995 		vrele(dvp);
2996 		return (0);
2997 	}
2998 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
2999 
3000 	mtx_unlock(vlp);
3001 	vn_lock(*vp, LK_SHARED | LK_RETRY);
3002 	error = VOP_VPTOCNP(*vp, &dvp, buf, buflen);
3003 	vput(*vp);
3004 	if (error) {
3005 		counter_u64_add(numfullpathfail2, 1);
3006 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
3007 		return (error);
3008 	}
3009 
3010 	*vp = dvp;
3011 	if (VN_IS_DOOMED(dvp)) {
3012 		/* forced unmount */
3013 		vrele(dvp);
3014 		error = ENOENT;
3015 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
3016 		return (error);
3017 	}
3018 	/*
3019 	 * *vp has its use count incremented still.
3020 	 */
3021 
3022 	return (0);
3023 }
3024 
3025 /*
3026  * Resolve a directory to a pathname.
3027  *
3028  * The name of the directory can always be found in the namecache or fetched
3029  * from the filesystem. There is also guaranteed to be only one parent, meaning
3030  * we can just follow vnodes up until we find the root.
3031  *
3032  * The vnode must be referenced.
3033  */
3034 static int
3035 vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3036     size_t *len, size_t addend)
3037 {
3038 #ifdef KDTRACE_HOOKS
3039 	struct vnode *startvp = vp;
3040 #endif
3041 	struct vnode *vp1;
3042 	size_t buflen;
3043 	int error;
3044 	bool slash_prefixed;
3045 
3046 	VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp);
3047 	VNPASS(vp->v_usecount > 0, vp);
3048 
3049 	buflen = *len;
3050 
3051 	slash_prefixed = true;
3052 	if (addend == 0) {
3053 		MPASS(*len >= 2);
3054 		buflen--;
3055 		buf[buflen] = '\0';
3056 		slash_prefixed = false;
3057 	}
3058 
3059 	error = 0;
3060 
3061 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
3062 	counter_u64_add(numfullpathcalls, 1);
3063 	while (vp != rdir && vp != rootvnode) {
3064 		/*
3065 		 * The vp vnode must be already fully constructed,
3066 		 * since it is either found in namecache or obtained
3067 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
3068 		 * without obtaining the vnode lock.
3069 		 */
3070 		if ((vp->v_vflag & VV_ROOT) != 0) {
3071 			vn_lock(vp, LK_RETRY | LK_SHARED);
3072 
3073 			/*
3074 			 * With the vnode locked, check for races with
3075 			 * unmount, forced or not.  Note that we
3076 			 * already verified that vp is not equal to
3077 			 * the root vnode, which means that
3078 			 * mnt_vnodecovered can be NULL only for the
3079 			 * case of unmount.
3080 			 */
3081 			if (VN_IS_DOOMED(vp) ||
3082 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
3083 			    vp1->v_mountedhere != vp->v_mount) {
3084 				vput(vp);
3085 				error = ENOENT;
3086 				SDT_PROBE3(vfs, namecache, fullpath, return,
3087 				    error, vp, NULL);
3088 				break;
3089 			}
3090 
3091 			vref(vp1);
3092 			vput(vp);
3093 			vp = vp1;
3094 			continue;
3095 		}
3096 		if (vp->v_type != VDIR) {
3097 			vrele(vp);
3098 			counter_u64_add(numfullpathfail1, 1);
3099 			error = ENOTDIR;
3100 			SDT_PROBE3(vfs, namecache, fullpath, return,
3101 			    error, vp, NULL);
3102 			break;
3103 		}
3104 		error = vn_vptocnp(&vp, buf, &buflen);
3105 		if (error)
3106 			break;
3107 		if (buflen == 0) {
3108 			vrele(vp);
3109 			error = ENOMEM;
3110 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
3111 			    startvp, NULL);
3112 			break;
3113 		}
3114 		buf[--buflen] = '/';
3115 		slash_prefixed = true;
3116 	}
3117 	if (error)
3118 		return (error);
3119 	if (!slash_prefixed) {
3120 		if (buflen == 0) {
3121 			vrele(vp);
3122 			counter_u64_add(numfullpathfail4, 1);
3123 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
3124 			    startvp, NULL);
3125 			return (ENOMEM);
3126 		}
3127 		buf[--buflen] = '/';
3128 	}
3129 	counter_u64_add(numfullpathfound, 1);
3130 	vrele(vp);
3131 
3132 	*retbuf = buf + buflen;
3133 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf);
3134 	*len -= buflen;
3135 	*len += addend;
3136 	return (0);
3137 }
3138 
3139 /*
3140  * Resolve an arbitrary vnode to a pathname.
3141  *
3142  * Note 2 caveats:
3143  * - hardlinks are not tracked, thus if the vnode is not a directory this can
3144  *   resolve to a different path than the one used to find it
3145  * - namecache is not mandatory, meaning names are not guaranteed to be added
3146  *   (in which case resolving fails)
3147  */
3148 static void __inline
3149 cache_rev_failed_impl(int *reason, int line)
3150 {
3151 
3152 	*reason = line;
3153 }
3154 #define cache_rev_failed(var)	cache_rev_failed_impl((var), __LINE__)
3155 
3156 static int
3157 vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
3158     char **retbuf, size_t *buflen, size_t addend)
3159 {
3160 #ifdef KDTRACE_HOOKS
3161 	struct vnode *startvp = vp;
3162 #endif
3163 	struct vnode *tvp;
3164 	struct mount *mp;
3165 	struct namecache *ncp;
3166 	size_t orig_buflen;
3167 	int reason;
3168 	int error;
3169 #ifdef KDTRACE_HOOKS
3170 	int i;
3171 #endif
3172 	seqc_t vp_seqc, tvp_seqc;
3173 	u_char nc_flag;
3174 
3175 	VFS_SMR_ASSERT_ENTERED();
3176 
3177 	if (!cache_fast_revlookup) {
3178 		vfs_smr_exit();
3179 		return (-1);
3180 	}
3181 
3182 	orig_buflen = *buflen;
3183 
3184 	if (addend == 0) {
3185 		MPASS(*buflen >= 2);
3186 		*buflen -= 1;
3187 		buf[*buflen] = '\0';
3188 	}
3189 
3190 	if (vp == rdir || vp == rootvnode) {
3191 		if (addend == 0) {
3192 			*buflen -= 1;
3193 			buf[*buflen] = '/';
3194 		}
3195 		goto out_ok;
3196 	}
3197 
3198 #ifdef KDTRACE_HOOKS
3199 	i = 0;
3200 #endif
3201 	error = -1;
3202 	ncp = NULL; /* for sdt probe down below */
3203 	vp_seqc = vn_seqc_read_any(vp);
3204 	if (seqc_in_modify(vp_seqc)) {
3205 		cache_rev_failed(&reason);
3206 		goto out_abort;
3207 	}
3208 
3209 	for (;;) {
3210 #ifdef KDTRACE_HOOKS
3211 		i++;
3212 #endif
3213 		if ((vp->v_vflag & VV_ROOT) != 0) {
3214 			mp = atomic_load_ptr(&vp->v_mount);
3215 			if (mp == NULL) {
3216 				cache_rev_failed(&reason);
3217 				goto out_abort;
3218 			}
3219 			tvp = atomic_load_ptr(&mp->mnt_vnodecovered);
3220 			tvp_seqc = vn_seqc_read_any(tvp);
3221 			if (seqc_in_modify(tvp_seqc)) {
3222 				cache_rev_failed(&reason);
3223 				goto out_abort;
3224 			}
3225 			if (!vn_seqc_consistent(vp, vp_seqc)) {
3226 				cache_rev_failed(&reason);
3227 				goto out_abort;
3228 			}
3229 			vp = tvp;
3230 			vp_seqc = tvp_seqc;
3231 			continue;
3232 		}
3233 		ncp = atomic_load_ptr(&vp->v_cache_dd);
3234 		if (ncp == NULL) {
3235 			cache_rev_failed(&reason);
3236 			goto out_abort;
3237 		}
3238 		nc_flag = atomic_load_char(&ncp->nc_flag);
3239 		if ((nc_flag & NCF_ISDOTDOT) != 0) {
3240 			cache_rev_failed(&reason);
3241 			goto out_abort;
3242 		}
3243 		if (!cache_ncp_canuse(ncp)) {
3244 			cache_rev_failed(&reason);
3245 			goto out_abort;
3246 		}
3247 		if (ncp->nc_nlen >= *buflen) {
3248 			cache_rev_failed(&reason);
3249 			error = ENOMEM;
3250 			goto out_abort;
3251 		}
3252 		*buflen -= ncp->nc_nlen;
3253 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
3254 		*buflen -= 1;
3255 		buf[*buflen] = '/';
3256 		tvp = ncp->nc_dvp;
3257 		tvp_seqc = vn_seqc_read_any(tvp);
3258 		if (seqc_in_modify(tvp_seqc)) {
3259 			cache_rev_failed(&reason);
3260 			goto out_abort;
3261 		}
3262 		if (!vn_seqc_consistent(vp, vp_seqc)) {
3263 			cache_rev_failed(&reason);
3264 			goto out_abort;
3265 		}
3266 		vp = tvp;
3267 		vp_seqc = tvp_seqc;
3268 		if (vp == rdir || vp == rootvnode)
3269 			break;
3270 	}
3271 out_ok:
3272 	vfs_smr_exit();
3273 	*retbuf = buf + *buflen;
3274 	*buflen = orig_buflen - *buflen + addend;
3275 	SDT_PROBE2(vfs, namecache, fullpath_smr, hit, startvp, *retbuf);
3276 	return (0);
3277 
3278 out_abort:
3279 	*buflen = orig_buflen;
3280 	SDT_PROBE4(vfs, namecache, fullpath_smr, miss, startvp, ncp, reason, i);
3281 	vfs_smr_exit();
3282 	return (error);
3283 }
3284 
3285 static int
3286 vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3287     size_t *buflen)
3288 {
3289 	size_t orig_buflen, addend;
3290 	int error;
3291 
3292 	if (*buflen < 2)
3293 		return (EINVAL);
3294 
3295 	orig_buflen = *buflen;
3296 
3297 	vref(vp);
3298 	addend = 0;
3299 	if (vp->v_type != VDIR) {
3300 		*buflen -= 1;
3301 		buf[*buflen] = '\0';
3302 		error = vn_vptocnp(&vp, buf, buflen);
3303 		if (error)
3304 			return (error);
3305 		if (*buflen == 0) {
3306 			vrele(vp);
3307 			return (ENOMEM);
3308 		}
3309 		*buflen -= 1;
3310 		buf[*buflen] = '/';
3311 		addend = orig_buflen - *buflen;
3312 	}
3313 
3314 	return (vn_fullpath_dir(vp, rdir, buf, retbuf, buflen, addend));
3315 }
3316 
3317 /*
3318  * Resolve an arbitrary vnode to a pathname (taking care of hardlinks).
3319  *
3320  * Since the namecache does not track hardlinks, the caller is expected to first
3321  * look up the target vnode with SAVENAME | WANTPARENT flags passed to namei.
3322  *
3323  * Then we have 2 cases:
3324  * - if the found vnode is a directory, the path can be constructed just by
3325  *   following names up the chain
3326  * - otherwise we populate the buffer with the saved name and start resolving
3327  *   from the parent
3328  */
3329 static int
3330 vn_fullpath_hardlink(struct nameidata *ndp, char **retbuf, char **freebuf,
3331     size_t *buflen)
3332 {
3333 	char *buf, *tmpbuf;
3334 	struct pwd *pwd;
3335 	struct componentname *cnp;
3336 	struct vnode *vp;
3337 	size_t addend;
3338 	int error;
3339 	enum vtype type;
3340 
3341 	if (*buflen < 2)
3342 		return (EINVAL);
3343 	if (*buflen > MAXPATHLEN)
3344 		*buflen = MAXPATHLEN;
3345 
3346 	buf = malloc(*buflen, M_TEMP, M_WAITOK);
3347 
3348 	addend = 0;
3349 	vp = ndp->ni_vp;
3350 	/*
3351 	 * Check for VBAD to work around the vp_crossmp bug in lookup().
3352 	 *
3353 	 * For example consider tmpfs on /tmp and realpath /tmp. ni_vp will be
3354 	 * set to mount point's root vnode while ni_dvp will be vp_crossmp.
3355 	 * If the type is VDIR (like in this very case) we can skip looking
3356 	 * at ni_dvp in the first place. However, since vnodes get passed here
3357 	 * unlocked the target may transition to doomed state (type == VBAD)
3358 	 * before we get to evaluate the condition. If this happens, we will
3359 	 * populate part of the buffer and descend to vn_fullpath_dir with
3360 	 * vp == vp_crossmp. Prevent the problem by checking for VBAD.
3361 	 *
3362 	 * This should be atomic_load(&vp->v_type) but it is illegal to take
3363 	 * an address of a bit field, even if said field is sized to char.
3364 	 * Work around the problem by reading the value into a full-sized enum
3365 	 * and then re-reading it with atomic_load which will still prevent
3366 	 * the compiler from re-reading down the road.
3367 	 */
3368 	type = vp->v_type;
3369 	type = atomic_load_int(&type);
3370 	if (type == VBAD) {
3371 		error = ENOENT;
3372 		goto out_bad;
3373 	}
3374 	if (type != VDIR) {
3375 		cnp = &ndp->ni_cnd;
3376 		addend = cnp->cn_namelen + 2;
3377 		if (*buflen < addend) {
3378 			error = ENOMEM;
3379 			goto out_bad;
3380 		}
3381 		*buflen -= addend;
3382 		tmpbuf = buf + *buflen;
3383 		tmpbuf[0] = '/';
3384 		memcpy(&tmpbuf[1], cnp->cn_nameptr, cnp->cn_namelen);
3385 		tmpbuf[addend - 1] = '\0';
3386 		vp = ndp->ni_dvp;
3387 	}
3388 
3389 	vfs_smr_enter();
3390 	pwd = pwd_get_smr();
3391 	error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3392 	    addend);
3393 	VFS_SMR_ASSERT_NOT_ENTERED();
3394 	if (error < 0) {
3395 		pwd = pwd_hold(curthread);
3396 		vref(vp);
3397 		error = vn_fullpath_dir(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3398 		    addend);
3399 		pwd_drop(pwd);
3400 		if (error != 0)
3401 			goto out_bad;
3402 	}
3403 
3404 	*freebuf = buf;
3405 
3406 	return (0);
3407 out_bad:
3408 	free(buf, M_TEMP);
3409 	return (error);
3410 }
3411 
3412 struct vnode *
3413 vn_dir_dd_ino(struct vnode *vp)
3414 {
3415 	struct namecache *ncp;
3416 	struct vnode *ddvp;
3417 	struct mtx *vlp;
3418 	enum vgetstate vs;
3419 
3420 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
3421 	vlp = VP2VNODELOCK(vp);
3422 	mtx_lock(vlp);
3423 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
3424 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
3425 			continue;
3426 		ddvp = ncp->nc_dvp;
3427 		vs = vget_prep(ddvp);
3428 		mtx_unlock(vlp);
3429 		if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
3430 			return (NULL);
3431 		return (ddvp);
3432 	}
3433 	mtx_unlock(vlp);
3434 	return (NULL);
3435 }
3436 
3437 int
3438 vn_commname(struct vnode *vp, char *buf, u_int buflen)
3439 {
3440 	struct namecache *ncp;
3441 	struct mtx *vlp;
3442 	int l;
3443 
3444 	vlp = VP2VNODELOCK(vp);
3445 	mtx_lock(vlp);
3446 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
3447 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
3448 			break;
3449 	if (ncp == NULL) {
3450 		mtx_unlock(vlp);
3451 		return (ENOENT);
3452 	}
3453 	l = min(ncp->nc_nlen, buflen - 1);
3454 	memcpy(buf, ncp->nc_name, l);
3455 	mtx_unlock(vlp);
3456 	buf[l] = '\0';
3457 	return (0);
3458 }
3459 
3460 /*
3461  * This function updates path string to vnode's full global path
3462  * and checks the size of the new path string against the pathlen argument.
3463  *
3464  * Requires a locked, referenced vnode.
3465  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
3466  *
3467  * If vp is a directory, the call to vn_fullpath_global() always succeeds
3468  * because it falls back to the ".." lookup if the namecache lookup fails.
3469  */
3470 int
3471 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
3472     u_int pathlen)
3473 {
3474 	struct nameidata nd;
3475 	struct vnode *vp1;
3476 	char *rpath, *fbuf;
3477 	int error;
3478 
3479 	ASSERT_VOP_ELOCKED(vp, __func__);
3480 
3481 	/* Construct global filesystem path from vp. */
3482 	VOP_UNLOCK(vp);
3483 	error = vn_fullpath_global(vp, &rpath, &fbuf);
3484 
3485 	if (error != 0) {
3486 		vrele(vp);
3487 		return (error);
3488 	}
3489 
3490 	if (strlen(rpath) >= pathlen) {
3491 		vrele(vp);
3492 		error = ENAMETOOLONG;
3493 		goto out;
3494 	}
3495 
3496 	/*
3497 	 * Re-lookup the vnode by path to detect a possible rename.
3498 	 * As a side effect, the vnode is relocked.
3499 	 * If vnode was renamed, return ENOENT.
3500 	 */
3501 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1,
3502 	    UIO_SYSSPACE, path, td);
3503 	error = namei(&nd);
3504 	if (error != 0) {
3505 		vrele(vp);
3506 		goto out;
3507 	}
3508 	NDFREE(&nd, NDF_ONLY_PNBUF);
3509 	vp1 = nd.ni_vp;
3510 	vrele(vp);
3511 	if (vp1 == vp)
3512 		strcpy(path, rpath);
3513 	else {
3514 		vput(vp1);
3515 		error = ENOENT;
3516 	}
3517 
3518 out:
3519 	free(fbuf, M_TEMP);
3520 	return (error);
3521 }
3522 
3523 #ifdef DDB
3524 static void
3525 db_print_vpath(struct vnode *vp)
3526 {
3527 
3528 	while (vp != NULL) {
3529 		db_printf("%p: ", vp);
3530 		if (vp == rootvnode) {
3531 			db_printf("/");
3532 			vp = NULL;
3533 		} else {
3534 			if (vp->v_vflag & VV_ROOT) {
3535 				db_printf("<mount point>");
3536 				vp = vp->v_mount->mnt_vnodecovered;
3537 			} else {
3538 				struct namecache *ncp;
3539 				char *ncn;
3540 				int i;
3541 
3542 				ncp = TAILQ_FIRST(&vp->v_cache_dst);
3543 				if (ncp != NULL) {
3544 					ncn = ncp->nc_name;
3545 					for (i = 0; i < ncp->nc_nlen; i++)
3546 						db_printf("%c", *ncn++);
3547 					vp = ncp->nc_dvp;
3548 				} else {
3549 					vp = NULL;
3550 				}
3551 			}
3552 		}
3553 		db_printf("\n");
3554 	}
3555 
3556 	return;
3557 }
3558 
3559 DB_SHOW_COMMAND(vpath, db_show_vpath)
3560 {
3561 	struct vnode *vp;
3562 
3563 	if (!have_addr) {
3564 		db_printf("usage: show vpath <struct vnode *>\n");
3565 		return;
3566 	}
3567 
3568 	vp = (struct vnode *)addr;
3569 	db_print_vpath(vp);
3570 }
3571 
3572 #endif
3573 
3574 static bool __read_frequently cache_fast_lookup = true;
3575 SYSCTL_BOOL(_vfs, OID_AUTO, cache_fast_lookup, CTLFLAG_RW,
3576     &cache_fast_lookup, 0, "");
3577 
3578 #define CACHE_FPL_FAILED	-2020
3579 
3580 /*
3581  * Components of nameidata (or objects it can point to) which may
3582  * need restoring in case fast path lookup fails.
3583  */
3584 struct nameidata_saved {
3585 	char *cn_nameptr;
3586 	size_t ni_pathlen;
3587 	int cn_flags;
3588 };
3589 
3590 struct cache_fpl {
3591 	struct nameidata *ndp;
3592 	struct componentname *cnp;
3593 	struct pwd *pwd;
3594 	struct vnode *dvp;
3595 	struct vnode *tvp;
3596 	seqc_t dvp_seqc;
3597 	seqc_t tvp_seqc;
3598 	struct nameidata_saved snd;
3599 	int line;
3600 	enum cache_fpl_status status:8;
3601 	bool in_smr;
3602 	bool fsearch;
3603 };
3604 
3605 static void
3606 cache_fpl_cleanup_cnp(struct componentname *cnp)
3607 {
3608 
3609 	uma_zfree(namei_zone, cnp->cn_pnbuf);
3610 #ifdef DIAGNOSTIC
3611 	cnp->cn_pnbuf = NULL;
3612 	cnp->cn_nameptr = NULL;
3613 #endif
3614 }
3615 
3616 static struct vnode *
3617 cache_fpl_handle_root(struct cache_fpl *fpl)
3618 {
3619 	struct nameidata *ndp;
3620 	struct componentname *cnp;
3621 
3622 	ndp = fpl->ndp;
3623 	cnp = fpl->cnp;
3624 
3625 	MPASS(*(cnp->cn_nameptr) == '/');
3626 	cnp->cn_nameptr++;
3627 	ndp->ni_pathlen--;
3628 
3629 	if (__predict_false(*(cnp->cn_nameptr) == '/')) {
3630 		do {
3631 			cnp->cn_nameptr++;
3632 			ndp->ni_pathlen--;
3633 		} while (*(cnp->cn_nameptr) == '/');
3634 	}
3635 
3636 	return (ndp->ni_rootdir);
3637 }
3638 
3639 static void
3640 cache_fpl_checkpoint(struct cache_fpl *fpl, struct nameidata_saved *snd)
3641 {
3642 
3643 	snd->cn_flags = fpl->ndp->ni_cnd.cn_flags;
3644 	snd->cn_nameptr = fpl->ndp->ni_cnd.cn_nameptr;
3645 	snd->ni_pathlen = fpl->ndp->ni_pathlen;
3646 }
3647 
3648 static void
3649 cache_fpl_restore_partial(struct cache_fpl *fpl, struct nameidata_saved *snd)
3650 {
3651 
3652 	fpl->ndp->ni_cnd.cn_flags = snd->cn_flags;
3653 	fpl->ndp->ni_cnd.cn_nameptr = snd->cn_nameptr;
3654 	fpl->ndp->ni_pathlen = snd->ni_pathlen;
3655 }
3656 
3657 static void
3658 cache_fpl_restore_abort(struct cache_fpl *fpl, struct nameidata_saved *snd)
3659 {
3660 
3661 	cache_fpl_restore_partial(fpl, snd);
3662 	/*
3663 	 * It is 0 on entry by API contract.
3664 	 */
3665 	fpl->ndp->ni_resflags = 0;
3666 }
3667 
3668 #ifdef INVARIANTS
3669 #define cache_fpl_smr_assert_entered(fpl) ({			\
3670 	struct cache_fpl *_fpl = (fpl);				\
3671 	MPASS(_fpl->in_smr == true);				\
3672 	VFS_SMR_ASSERT_ENTERED();				\
3673 })
3674 #define cache_fpl_smr_assert_not_entered(fpl) ({		\
3675 	struct cache_fpl *_fpl = (fpl);				\
3676 	MPASS(_fpl->in_smr == false);				\
3677 	VFS_SMR_ASSERT_NOT_ENTERED();				\
3678 })
3679 #else
3680 #define cache_fpl_smr_assert_entered(fpl) do { } while (0)
3681 #define cache_fpl_smr_assert_not_entered(fpl) do { } while (0)
3682 #endif
3683 
3684 #define cache_fpl_smr_enter_initial(fpl) ({			\
3685 	struct cache_fpl *_fpl = (fpl);				\
3686 	vfs_smr_enter();					\
3687 	_fpl->in_smr = true;					\
3688 })
3689 
3690 #define cache_fpl_smr_enter(fpl) ({				\
3691 	struct cache_fpl *_fpl = (fpl);				\
3692 	MPASS(_fpl->in_smr == false);				\
3693 	vfs_smr_enter();					\
3694 	_fpl->in_smr = true;					\
3695 })
3696 
3697 #define cache_fpl_smr_exit(fpl) ({				\
3698 	struct cache_fpl *_fpl = (fpl);				\
3699 	MPASS(_fpl->in_smr == true);				\
3700 	vfs_smr_exit();						\
3701 	_fpl->in_smr = false;					\
3702 })
3703 
3704 static int
3705 cache_fpl_aborted_impl(struct cache_fpl *fpl, int line)
3706 {
3707 
3708 	if (fpl->status != CACHE_FPL_STATUS_UNSET) {
3709 		KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL,
3710 		    ("%s: converting to abort from %d at %d, set at %d\n",
3711 		    __func__, fpl->status, line, fpl->line));
3712 	}
3713 	fpl->status = CACHE_FPL_STATUS_ABORTED;
3714 	fpl->line = line;
3715 	return (CACHE_FPL_FAILED);
3716 }
3717 
3718 #define cache_fpl_aborted(x)	cache_fpl_aborted_impl((x), __LINE__)
3719 
3720 static int
3721 cache_fpl_partial_impl(struct cache_fpl *fpl, int line)
3722 {
3723 
3724 	KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3725 	    ("%s: setting to partial at %d, but already set to %d at %d\n",
3726 	    __func__, line, fpl->status, fpl->line));
3727 	cache_fpl_smr_assert_entered(fpl);
3728 	fpl->status = CACHE_FPL_STATUS_PARTIAL;
3729 	fpl->line = line;
3730 	return (CACHE_FPL_FAILED);
3731 }
3732 
3733 #define cache_fpl_partial(x)	cache_fpl_partial_impl((x), __LINE__)
3734 
3735 static int
3736 cache_fpl_handled_impl(struct cache_fpl *fpl, int error, int line)
3737 {
3738 
3739 	KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
3740 	    ("%s: setting to handled at %d, but already set to %d at %d\n",
3741 	    __func__, line, fpl->status, fpl->line));
3742 	cache_fpl_smr_assert_not_entered(fpl);
3743 	MPASS(error != CACHE_FPL_FAILED);
3744 	fpl->status = CACHE_FPL_STATUS_HANDLED;
3745 	fpl->line = line;
3746 	return (error);
3747 }
3748 
3749 #define cache_fpl_handled(x, e)	cache_fpl_handled_impl((x), (e), __LINE__)
3750 
3751 static bool
3752 cache_fpl_terminated(struct cache_fpl *fpl)
3753 {
3754 
3755 	return (fpl->status != CACHE_FPL_STATUS_UNSET);
3756 }
3757 
3758 #define CACHE_FPL_SUPPORTED_CN_FLAGS \
3759 	(NC_NOMAKEENTRY | NC_KEEPPOSENTRY | LOCKLEAF | LOCKPARENT | WANTPARENT | \
3760 	 FAILIFEXISTS | FOLLOW | LOCKSHARED | SAVENAME | SAVESTART | WILLBEDIR | \
3761 	 ISOPEN | NOMACCHECK | AUDITVNODE1 | AUDITVNODE2 | NOCAPCHECK)
3762 
3763 #define CACHE_FPL_INTERNAL_CN_FLAGS \
3764 	(ISDOTDOT | MAKEENTRY | ISLASTCN)
3765 
3766 _Static_assert((CACHE_FPL_SUPPORTED_CN_FLAGS & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
3767     "supported and internal flags overlap");
3768 
3769 static bool cache_fplookup_is_mp(struct cache_fpl *fpl);
3770 static int cache_fplookup_cross_mount(struct cache_fpl *fpl);
3771 
3772 static bool
3773 cache_fpl_islastcn(struct nameidata *ndp)
3774 {
3775 
3776 	return (*ndp->ni_next == 0);
3777 }
3778 
3779 static bool
3780 cache_fpl_isdotdot(struct componentname *cnp)
3781 {
3782 
3783 	if (cnp->cn_namelen == 2 &&
3784 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
3785 		return (true);
3786 	return (false);
3787 }
3788 
3789 static bool
3790 cache_can_fplookup(struct cache_fpl *fpl)
3791 {
3792 	struct nameidata *ndp;
3793 	struct componentname *cnp;
3794 	struct thread *td;
3795 
3796 	ndp = fpl->ndp;
3797 	cnp = fpl->cnp;
3798 	td = cnp->cn_thread;
3799 
3800 	if (!cache_fast_lookup) {
3801 		cache_fpl_aborted(fpl);
3802 		return (false);
3803 	}
3804 #ifdef MAC
3805 	if (mac_vnode_check_lookup_enabled()) {
3806 		cache_fpl_aborted(fpl);
3807 		return (false);
3808 	}
3809 #endif
3810 	if ((cnp->cn_flags & ~CACHE_FPL_SUPPORTED_CN_FLAGS) != 0) {
3811 		cache_fpl_aborted(fpl);
3812 		return (false);
3813 	}
3814 	if (IN_CAPABILITY_MODE(td)) {
3815 		cache_fpl_aborted(fpl);
3816 		return (false);
3817 	}
3818 	if (AUDITING_TD(td)) {
3819 		cache_fpl_aborted(fpl);
3820 		return (false);
3821 	}
3822 	if (ndp->ni_startdir != NULL) {
3823 		cache_fpl_aborted(fpl);
3824 		return (false);
3825 	}
3826 	return (true);
3827 }
3828 
3829 static int
3830 cache_fplookup_dirfd(struct cache_fpl *fpl, struct vnode **vpp)
3831 {
3832 	struct nameidata *ndp;
3833 	int error;
3834 	bool fsearch;
3835 
3836 	ndp = fpl->ndp;
3837 	error = fgetvp_lookup_smr(ndp->ni_dirfd, ndp, vpp, &fsearch);
3838 	if (__predict_false(error != 0)) {
3839 		cache_fpl_smr_exit(fpl);
3840 		return (cache_fpl_aborted(fpl));
3841 	}
3842 	fpl->fsearch = fsearch;
3843 	return (0);
3844 }
3845 
3846 static bool
3847 cache_fplookup_vnode_supported(struct vnode *vp)
3848 {
3849 
3850 	return (vp->v_type != VLNK);
3851 }
3852 
3853 static int __noinline
3854 cache_fplookup_negative_promote(struct cache_fpl *fpl, struct namecache *oncp,
3855     uint32_t hash)
3856 {
3857 	struct componentname *cnp;
3858 	struct vnode *dvp;
3859 
3860 	cnp = fpl->cnp;
3861 	dvp = fpl->dvp;
3862 
3863 	cache_fpl_smr_exit(fpl);
3864 	if (cache_neg_promote_cond(dvp, cnp, oncp, hash))
3865 		return (cache_fpl_handled(fpl, ENOENT));
3866 	else
3867 		return (cache_fpl_aborted(fpl));
3868 }
3869 
3870 /*
3871  * The target vnode is not supported, prepare for the slow path to take over.
3872  */
3873 static int __noinline
3874 cache_fplookup_partial_setup(struct cache_fpl *fpl)
3875 {
3876 	struct nameidata *ndp;
3877 	struct componentname *cnp;
3878 	enum vgetstate dvs;
3879 	struct vnode *dvp;
3880 	struct pwd *pwd;
3881 	seqc_t dvp_seqc;
3882 
3883 	ndp = fpl->ndp;
3884 	cnp = fpl->cnp;
3885 	pwd = fpl->pwd;
3886 	dvp = fpl->dvp;
3887 	dvp_seqc = fpl->dvp_seqc;
3888 
3889 	if (!pwd_hold_smr(pwd)) {
3890 		cache_fpl_smr_exit(fpl);
3891 		return (cache_fpl_aborted(fpl));
3892 	}
3893 
3894 	/*
3895 	 * Note that seqc is checked before the vnode is locked, so by
3896 	 * the time regular lookup gets to it it may have moved.
3897 	 *
3898 	 * Ultimately this does not affect correctness, any lookup errors
3899 	 * are userspace racing with itself. It is guaranteed that any
3900 	 * path which ultimatley gets found could also have been found
3901 	 * by regular lookup going all the way in absence of concurrent
3902 	 * modifications.
3903 	 */
3904 	dvs = vget_prep_smr(dvp);
3905 	cache_fpl_smr_exit(fpl);
3906 	if (__predict_false(dvs == VGET_NONE)) {
3907 		pwd_drop(pwd);
3908 		return (cache_fpl_aborted(fpl));
3909 	}
3910 
3911 	vget_finish_ref(dvp, dvs);
3912 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
3913 		vrele(dvp);
3914 		pwd_drop(pwd);
3915 		return (cache_fpl_aborted(fpl));
3916 	}
3917 
3918 	cache_fpl_restore_partial(fpl, &fpl->snd);
3919 
3920 	ndp->ni_startdir = dvp;
3921 	cnp->cn_flags |= MAKEENTRY;
3922 	if (cache_fpl_islastcn(ndp))
3923 		cnp->cn_flags |= ISLASTCN;
3924 	if (cache_fpl_isdotdot(cnp))
3925 		cnp->cn_flags |= ISDOTDOT;
3926 
3927 	return (0);
3928 }
3929 
3930 static int
3931 cache_fplookup_final_child(struct cache_fpl *fpl, enum vgetstate tvs)
3932 {
3933 	struct componentname *cnp;
3934 	struct vnode *tvp;
3935 	seqc_t tvp_seqc;
3936 	int error, lkflags;
3937 
3938 	cnp = fpl->cnp;
3939 	tvp = fpl->tvp;
3940 	tvp_seqc = fpl->tvp_seqc;
3941 
3942 	if ((cnp->cn_flags & LOCKLEAF) != 0) {
3943 		lkflags = LK_SHARED;
3944 		if ((cnp->cn_flags & LOCKSHARED) == 0)
3945 			lkflags = LK_EXCLUSIVE;
3946 		error = vget_finish(tvp, lkflags, tvs);
3947 		if (__predict_false(error != 0)) {
3948 			return (cache_fpl_aborted(fpl));
3949 		}
3950 	} else {
3951 		vget_finish_ref(tvp, tvs);
3952 	}
3953 
3954 	if (!vn_seqc_consistent(tvp, tvp_seqc)) {
3955 		if ((cnp->cn_flags & LOCKLEAF) != 0)
3956 			vput(tvp);
3957 		else
3958 			vrele(tvp);
3959 		return (cache_fpl_aborted(fpl));
3960 	}
3961 
3962 	return (cache_fpl_handled(fpl, 0));
3963 }
3964 
3965 /*
3966  * They want to possibly modify the state of the namecache.
3967  */
3968 static int __noinline
3969 cache_fplookup_final_modifying(struct cache_fpl *fpl)
3970 {
3971 	struct nameidata *ndp;
3972 	struct componentname *cnp;
3973 	enum vgetstate dvs;
3974 	struct vnode *dvp, *tvp;
3975 	struct mount *mp;
3976 	seqc_t dvp_seqc;
3977 	int error;
3978 	bool docache;
3979 
3980 	ndp = fpl->ndp;
3981 	cnp = fpl->cnp;
3982 	dvp = fpl->dvp;
3983 	dvp_seqc = fpl->dvp_seqc;
3984 
3985 	MPASS(cache_fpl_islastcn(ndp));
3986 	if ((cnp->cn_flags & LOCKPARENT) == 0)
3987 		MPASS((cnp->cn_flags & WANTPARENT) != 0);
3988 	MPASS((cnp->cn_flags & TRAILINGSLASH) == 0);
3989 	MPASS(cnp->cn_nameiop == CREATE || cnp->cn_nameiop == DELETE ||
3990 	    cnp->cn_nameiop == RENAME);
3991 	MPASS((cnp->cn_flags & MAKEENTRY) == 0);
3992 	MPASS((cnp->cn_flags & ISDOTDOT) == 0);
3993 
3994 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
3995 	if (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)
3996 		docache = false;
3997 
3998 	mp = atomic_load_ptr(&dvp->v_mount);
3999 	if (__predict_false(mp == NULL)) {
4000 		return (cache_fpl_aborted(fpl));
4001 	}
4002 
4003 	if (__predict_false(mp->mnt_flag & MNT_RDONLY)) {
4004 		cache_fpl_smr_exit(fpl);
4005 		/*
4006 		 * Original code keeps not checking for CREATE which
4007 		 * might be a bug. For now let the old lookup decide.
4008 		 */
4009 		if (cnp->cn_nameiop == CREATE) {
4010 			return (cache_fpl_aborted(fpl));
4011 		}
4012 		return (cache_fpl_handled(fpl, EROFS));
4013 	}
4014 
4015 	if (fpl->tvp != NULL && (cnp->cn_flags & FAILIFEXISTS) != 0) {
4016 		cache_fpl_smr_exit(fpl);
4017 		return (cache_fpl_handled(fpl, EEXIST));
4018 	}
4019 
4020 	/*
4021 	 * Secure access to dvp; check cache_fplookup_partial_setup for
4022 	 * reasoning.
4023 	 *
4024 	 * XXX At least UFS requires its lookup routine to be called for
4025 	 * the last path component, which leads to some level of complicaton
4026 	 * and inefficiency:
4027 	 * - the target routine always locks the target vnode, but our caller
4028 	 *   may not need it locked
4029 	 * - some of the VOP machinery asserts that the parent is locked, which
4030 	 *   once more may be not required
4031 	 *
4032 	 * TODO: add a flag for filesystems which don't need this.
4033 	 */
4034 	dvs = vget_prep_smr(dvp);
4035 	cache_fpl_smr_exit(fpl);
4036 	if (__predict_false(dvs == VGET_NONE)) {
4037 		return (cache_fpl_aborted(fpl));
4038 	}
4039 
4040 	vget_finish_ref(dvp, dvs);
4041 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4042 		vrele(dvp);
4043 		return (cache_fpl_aborted(fpl));
4044 	}
4045 
4046 	error = vn_lock(dvp, LK_EXCLUSIVE);
4047 	if (__predict_false(error != 0)) {
4048 		vrele(dvp);
4049 		return (cache_fpl_aborted(fpl));
4050 	}
4051 
4052 	tvp = NULL;
4053 	cnp->cn_flags |= ISLASTCN;
4054 	if (docache)
4055 		cnp->cn_flags |= MAKEENTRY;
4056 	if (cache_fpl_isdotdot(cnp))
4057 		cnp->cn_flags |= ISDOTDOT;
4058 	cnp->cn_lkflags = LK_EXCLUSIVE;
4059 	error = VOP_LOOKUP(dvp, &tvp, cnp);
4060 	switch (error) {
4061 	case EJUSTRETURN:
4062 	case 0:
4063 		break;
4064 	case ENOTDIR:
4065 	case ENOENT:
4066 		vput(dvp);
4067 		return (cache_fpl_handled(fpl, error));
4068 	default:
4069 		vput(dvp);
4070 		return (cache_fpl_aborted(fpl));
4071 	}
4072 
4073 	fpl->tvp = tvp;
4074 
4075 	if (tvp == NULL) {
4076 		if ((cnp->cn_flags & SAVESTART) != 0) {
4077 			ndp->ni_startdir = dvp;
4078 			vrefact(ndp->ni_startdir);
4079 			cnp->cn_flags |= SAVENAME;
4080 		}
4081 		MPASS(error == EJUSTRETURN);
4082 		if ((cnp->cn_flags & LOCKPARENT) == 0) {
4083 			VOP_UNLOCK(dvp);
4084 		}
4085 		return (cache_fpl_handled(fpl, 0));
4086 	}
4087 
4088 	/*
4089 	 * There are very hairy corner cases concerning various flag combinations
4090 	 * and locking state. In particular here we only hold one lock instead of
4091 	 * two.
4092 	 *
4093 	 * Skip the complexity as it is of no significance for normal workloads.
4094 	 */
4095 	if (__predict_false(tvp == dvp)) {
4096 		vput(dvp);
4097 		vrele(tvp);
4098 		return (cache_fpl_aborted(fpl));
4099 	}
4100 
4101 	/*
4102 	 * Check if the target is either a symlink or a mount point.
4103 	 * Since we expect this to be the terminal vnode it should
4104 	 * almost never be true.
4105 	 */
4106 	if (__predict_false(!cache_fplookup_vnode_supported(tvp) ||
4107 	    cache_fplookup_is_mp(fpl))) {
4108 		vput(dvp);
4109 		vput(tvp);
4110 		return (cache_fpl_aborted(fpl));
4111 	}
4112 
4113 	if ((cnp->cn_flags & FAILIFEXISTS) != 0) {
4114 		vput(dvp);
4115 		vput(tvp);
4116 		return (cache_fpl_handled(fpl, EEXIST));
4117 	}
4118 
4119 	if ((cnp->cn_flags & LOCKLEAF) == 0) {
4120 		VOP_UNLOCK(tvp);
4121 	}
4122 
4123 	if ((cnp->cn_flags & LOCKPARENT) == 0) {
4124 		VOP_UNLOCK(dvp);
4125 	}
4126 
4127 	if ((cnp->cn_flags & SAVESTART) != 0) {
4128 		ndp->ni_startdir = dvp;
4129 		vrefact(ndp->ni_startdir);
4130 		cnp->cn_flags |= SAVENAME;
4131 	}
4132 
4133 	return (cache_fpl_handled(fpl, 0));
4134 }
4135 
4136 static int __noinline
4137 cache_fplookup_modifying(struct cache_fpl *fpl)
4138 {
4139 	struct nameidata *ndp;
4140 
4141 	ndp = fpl->ndp;
4142 
4143 	if (!cache_fpl_islastcn(ndp)) {
4144 		return (cache_fpl_partial(fpl));
4145 	}
4146 	return  (cache_fplookup_final_modifying(fpl));
4147 }
4148 
4149 static int __noinline
4150 cache_fplookup_final_withparent(struct cache_fpl *fpl)
4151 {
4152 	struct componentname *cnp;
4153 	enum vgetstate dvs, tvs;
4154 	struct vnode *dvp, *tvp;
4155 	seqc_t dvp_seqc;
4156 	int error;
4157 
4158 	cnp = fpl->cnp;
4159 	dvp = fpl->dvp;
4160 	dvp_seqc = fpl->dvp_seqc;
4161 	tvp = fpl->tvp;
4162 
4163 	MPASS((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0);
4164 
4165 	/*
4166 	 * This is less efficient than it can be for simplicity.
4167 	 */
4168 	dvs = vget_prep_smr(dvp);
4169 	if (__predict_false(dvs == VGET_NONE)) {
4170 		return (cache_fpl_aborted(fpl));
4171 	}
4172 	tvs = vget_prep_smr(tvp);
4173 	if (__predict_false(tvs == VGET_NONE)) {
4174 		cache_fpl_smr_exit(fpl);
4175 		vget_abort(dvp, dvs);
4176 		return (cache_fpl_aborted(fpl));
4177 	}
4178 
4179 	cache_fpl_smr_exit(fpl);
4180 
4181 	if ((cnp->cn_flags & LOCKPARENT) != 0) {
4182 		error = vget_finish(dvp, LK_EXCLUSIVE, dvs);
4183 		if (__predict_false(error != 0)) {
4184 			vget_abort(tvp, tvs);
4185 			return (cache_fpl_aborted(fpl));
4186 		}
4187 	} else {
4188 		vget_finish_ref(dvp, dvs);
4189 	}
4190 
4191 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4192 		vget_abort(tvp, tvs);
4193 		if ((cnp->cn_flags & LOCKPARENT) != 0)
4194 			vput(dvp);
4195 		else
4196 			vrele(dvp);
4197 		return (cache_fpl_aborted(fpl));
4198 	}
4199 
4200 	error = cache_fplookup_final_child(fpl, tvs);
4201 	if (__predict_false(error != 0)) {
4202 		MPASS(fpl->status == CACHE_FPL_STATUS_ABORTED);
4203 		if ((cnp->cn_flags & LOCKPARENT) != 0)
4204 			vput(dvp);
4205 		else
4206 			vrele(dvp);
4207 		return (error);
4208 	}
4209 
4210 	MPASS(fpl->status == CACHE_FPL_STATUS_HANDLED);
4211 	return (0);
4212 }
4213 
4214 static int
4215 cache_fplookup_final(struct cache_fpl *fpl)
4216 {
4217 	struct componentname *cnp;
4218 	enum vgetstate tvs;
4219 	struct vnode *dvp, *tvp;
4220 	seqc_t dvp_seqc;
4221 
4222 	cnp = fpl->cnp;
4223 	dvp = fpl->dvp;
4224 	dvp_seqc = fpl->dvp_seqc;
4225 	tvp = fpl->tvp;
4226 
4227 	if (cnp->cn_nameiop != LOOKUP) {
4228 		return (cache_fplookup_final_modifying(fpl));
4229 	}
4230 
4231 	if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0)
4232 		return (cache_fplookup_final_withparent(fpl));
4233 
4234 	tvs = vget_prep_smr(tvp);
4235 	if (__predict_false(tvs == VGET_NONE)) {
4236 		return (cache_fpl_partial(fpl));
4237 	}
4238 
4239 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4240 		cache_fpl_smr_exit(fpl);
4241 		vget_abort(tvp, tvs);
4242 		return (cache_fpl_aborted(fpl));
4243 	}
4244 
4245 	cache_fpl_smr_exit(fpl);
4246 	return (cache_fplookup_final_child(fpl, tvs));
4247 }
4248 
4249 /*
4250  * Comment from locked lookup:
4251  * Check for degenerate name (e.g. / or "") which is a way of talking about a
4252  * directory, e.g. like "/." or ".".
4253  */
4254 static int __noinline
4255 cache_fplookup_degenerate(struct cache_fpl *fpl)
4256 {
4257 	struct componentname *cnp;
4258 	struct vnode *dvp;
4259 	enum vgetstate dvs;
4260 	int error, lkflags;
4261 
4262 	fpl->tvp = fpl->dvp;
4263 	fpl->tvp_seqc = fpl->dvp_seqc;
4264 
4265 	cnp = fpl->cnp;
4266 	dvp = fpl->dvp;
4267 
4268 	if (__predict_false(cnp->cn_nameiop != LOOKUP)) {
4269 		cache_fpl_smr_exit(fpl);
4270 		return (cache_fpl_handled(fpl, EISDIR));
4271 	}
4272 
4273 	MPASS((cnp->cn_flags & SAVESTART) == 0);
4274 
4275 	if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0) {
4276 		return (cache_fplookup_final_withparent(fpl));
4277 	}
4278 
4279 	dvs = vget_prep_smr(dvp);
4280 	cache_fpl_smr_exit(fpl);
4281 	if (__predict_false(dvs == VGET_NONE)) {
4282 		return (cache_fpl_aborted(fpl));
4283 	}
4284 
4285 	if ((cnp->cn_flags & LOCKLEAF) != 0) {
4286 		lkflags = LK_SHARED;
4287 		if ((cnp->cn_flags & LOCKSHARED) == 0)
4288 			lkflags = LK_EXCLUSIVE;
4289 		error = vget_finish(dvp, lkflags, dvs);
4290 		if (__predict_false(error != 0)) {
4291 			return (cache_fpl_aborted(fpl));
4292 		}
4293 	} else {
4294 		vget_finish_ref(dvp, dvs);
4295 	}
4296 	return (cache_fpl_handled(fpl, 0));
4297 }
4298 
4299 static int __noinline
4300 cache_fplookup_noentry(struct cache_fpl *fpl)
4301 {
4302 	struct nameidata *ndp;
4303 	struct componentname *cnp;
4304 	enum vgetstate dvs;
4305 	struct vnode *dvp, *tvp;
4306 	seqc_t dvp_seqc;
4307 	int error;
4308 	bool docache;
4309 
4310 	ndp = fpl->ndp;
4311 	cnp = fpl->cnp;
4312 	dvp = fpl->dvp;
4313 	dvp_seqc = fpl->dvp_seqc;
4314 
4315 	MPASS((cnp->cn_flags & MAKEENTRY) == 0);
4316 	MPASS((cnp->cn_flags & ISDOTDOT) == 0);
4317 	MPASS(!cache_fpl_isdotdot(cnp));
4318 
4319 	/*
4320 	 * Hack: delayed name len checking.
4321 	 */
4322 	if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
4323 		cache_fpl_smr_exit(fpl);
4324 		return (cache_fpl_handled(fpl, ENAMETOOLONG));
4325 	}
4326 
4327 	if (cnp->cn_nameiop != LOOKUP) {
4328 		fpl->tvp = NULL;
4329 		return (cache_fplookup_modifying(fpl));
4330 	}
4331 
4332 	MPASS((cnp->cn_flags & SAVESTART) == 0);
4333 
4334 	/*
4335 	 * Only try to fill in the component if it is the last one,
4336 	 * otherwise not only there may be several to handle but the
4337 	 * walk may be complicated.
4338 	 */
4339 	if (!cache_fpl_islastcn(ndp)) {
4340 		return (cache_fpl_partial(fpl));
4341 	}
4342 
4343 	/*
4344 	 * Secure access to dvp; check cache_fplookup_partial_setup for
4345 	 * reasoning.
4346 	 */
4347 	dvs = vget_prep_smr(dvp);
4348 	cache_fpl_smr_exit(fpl);
4349 	if (__predict_false(dvs == VGET_NONE)) {
4350 		return (cache_fpl_aborted(fpl));
4351 	}
4352 
4353 	vget_finish_ref(dvp, dvs);
4354 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4355 		vrele(dvp);
4356 		return (cache_fpl_aborted(fpl));
4357 	}
4358 
4359 	error = vn_lock(dvp, LK_SHARED);
4360 	if (__predict_false(error != 0)) {
4361 		vrele(dvp);
4362 		return (cache_fpl_aborted(fpl));
4363 	}
4364 
4365 	tvp = NULL;
4366 	/*
4367 	 * TODO: provide variants which don't require locking either vnode.
4368 	 */
4369 	cnp->cn_flags |= ISLASTCN;
4370 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
4371 	if (docache)
4372 		cnp->cn_flags |= MAKEENTRY;
4373 	cnp->cn_lkflags = LK_SHARED;
4374 	if ((cnp->cn_flags & LOCKSHARED) == 0) {
4375 		cnp->cn_lkflags = LK_EXCLUSIVE;
4376 	}
4377 	error = VOP_LOOKUP(dvp, &tvp, cnp);
4378 	switch (error) {
4379 	case EJUSTRETURN:
4380 	case 0:
4381 		break;
4382 	case ENOTDIR:
4383 	case ENOENT:
4384 		vput(dvp);
4385 		return (cache_fpl_handled(fpl, error));
4386 	default:
4387 		vput(dvp);
4388 		return (cache_fpl_aborted(fpl));
4389 	}
4390 
4391 	fpl->tvp = tvp;
4392 
4393 	if (tvp == NULL) {
4394 		MPASS(error == EJUSTRETURN);
4395 		if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) {
4396 			vput(dvp);
4397 		} else if ((cnp->cn_flags & LOCKPARENT) == 0) {
4398 			VOP_UNLOCK(dvp);
4399 		}
4400 		return (cache_fpl_handled(fpl, 0));
4401 	}
4402 
4403 	if (__predict_false(!cache_fplookup_vnode_supported(tvp) ||
4404 	    cache_fplookup_is_mp(fpl))) {
4405 		vput(dvp);
4406 		vput(tvp);
4407 		return (cache_fpl_aborted(fpl));
4408 	}
4409 
4410 	if ((cnp->cn_flags & LOCKLEAF) == 0) {
4411 		VOP_UNLOCK(tvp);
4412 	}
4413 
4414 	if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) {
4415 		vput(dvp);
4416 	} else if ((cnp->cn_flags & LOCKPARENT) == 0) {
4417 		VOP_UNLOCK(dvp);
4418 	}
4419 	return (cache_fpl_handled(fpl, 0));
4420 }
4421 
4422 static int __noinline
4423 cache_fplookup_dot(struct cache_fpl *fpl)
4424 {
4425 	int error;
4426 
4427 	MPASS(!seqc_in_modify(fpl->dvp_seqc));
4428 	/*
4429 	 * Just re-assign the value. seqc will be checked later for the first
4430 	 * non-dot path component in line and/or before deciding to return the
4431 	 * vnode.
4432 	 */
4433 	fpl->tvp = fpl->dvp;
4434 	fpl->tvp_seqc = fpl->dvp_seqc;
4435 	if (cache_fplookup_is_mp(fpl)) {
4436 		error = cache_fplookup_cross_mount(fpl);
4437 		if (__predict_false(error != 0)) {
4438 			return (error);
4439 		}
4440 	}
4441 
4442 	counter_u64_add(dothits, 1);
4443 	SDT_PROBE3(vfs, namecache, lookup, hit, fpl->dvp, ".", fpl->dvp);
4444 	return (0);
4445 }
4446 
4447 static int __noinline
4448 cache_fplookup_dotdot(struct cache_fpl *fpl)
4449 {
4450 	struct nameidata *ndp;
4451 	struct componentname *cnp;
4452 	struct namecache *ncp;
4453 	struct vnode *dvp;
4454 	struct prison *pr;
4455 	u_char nc_flag;
4456 
4457 	ndp = fpl->ndp;
4458 	cnp = fpl->cnp;
4459 	dvp = fpl->dvp;
4460 
4461 	MPASS(cache_fpl_isdotdot(cnp));
4462 
4463 	/*
4464 	 * XXX this is racy the same way regular lookup is
4465 	 */
4466 	for (pr = cnp->cn_cred->cr_prison; pr != NULL;
4467 	    pr = pr->pr_parent)
4468 		if (dvp == pr->pr_root)
4469 			break;
4470 
4471 	if (dvp == ndp->ni_rootdir ||
4472 	    dvp == ndp->ni_topdir ||
4473 	    dvp == rootvnode ||
4474 	    pr != NULL) {
4475 		fpl->tvp = dvp;
4476 		fpl->tvp_seqc = vn_seqc_read_any(dvp);
4477 		if (seqc_in_modify(fpl->tvp_seqc)) {
4478 			return (cache_fpl_aborted(fpl));
4479 		}
4480 		return (0);
4481 	}
4482 
4483 	if ((dvp->v_vflag & VV_ROOT) != 0) {
4484 		/*
4485 		 * TODO
4486 		 * The opposite of climb mount is needed here.
4487 		 */
4488 		return (cache_fpl_aborted(fpl));
4489 	}
4490 
4491 	ncp = atomic_load_ptr(&dvp->v_cache_dd);
4492 	if (ncp == NULL) {
4493 		return (cache_fpl_aborted(fpl));
4494 	}
4495 
4496 	nc_flag = atomic_load_char(&ncp->nc_flag);
4497 	if ((nc_flag & NCF_ISDOTDOT) != 0) {
4498 		if ((nc_flag & NCF_NEGATIVE) != 0)
4499 			return (cache_fpl_aborted(fpl));
4500 		fpl->tvp = ncp->nc_vp;
4501 	} else {
4502 		fpl->tvp = ncp->nc_dvp;
4503 	}
4504 
4505 	if (!cache_ncp_canuse(ncp)) {
4506 		return (cache_fpl_aborted(fpl));
4507 	}
4508 
4509 	fpl->tvp_seqc = vn_seqc_read_any(fpl->tvp);
4510 	if (seqc_in_modify(fpl->tvp_seqc)) {
4511 		return (cache_fpl_partial(fpl));
4512 	}
4513 
4514 	counter_u64_add(dotdothits, 1);
4515 	return (0);
4516 }
4517 
4518 static int __noinline
4519 cache_fplookup_neg(struct cache_fpl *fpl, struct namecache *ncp, uint32_t hash)
4520 {
4521 	u_char nc_flag;
4522 	bool neg_promote;
4523 
4524 	nc_flag = atomic_load_char(&ncp->nc_flag);
4525 	MPASS((nc_flag & NCF_NEGATIVE) != 0);
4526 	/*
4527 	 * If they want to create an entry we need to replace this one.
4528 	 */
4529 	if (__predict_false(fpl->cnp->cn_nameiop != LOOKUP)) {
4530 		fpl->tvp = NULL;
4531 		return (cache_fplookup_modifying(fpl));
4532 	}
4533 	neg_promote = cache_neg_hit_prep(ncp);
4534 	if (!cache_fpl_neg_ncp_canuse(ncp)) {
4535 		cache_neg_hit_abort(ncp);
4536 		return (cache_fpl_partial(fpl));
4537 	}
4538 	if (neg_promote) {
4539 		return (cache_fplookup_negative_promote(fpl, ncp, hash));
4540 	}
4541 	cache_neg_hit_finish(ncp);
4542 	cache_fpl_smr_exit(fpl);
4543 	return (cache_fpl_handled(fpl, ENOENT));
4544 }
4545 
4546 static int
4547 cache_fplookup_next(struct cache_fpl *fpl)
4548 {
4549 	struct componentname *cnp;
4550 	struct namecache *ncp;
4551 	struct vnode *dvp, *tvp;
4552 	u_char nc_flag;
4553 	uint32_t hash;
4554 	int error;
4555 
4556 	cnp = fpl->cnp;
4557 	dvp = fpl->dvp;
4558 
4559 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
4560 		if (cnp->cn_namelen == 1) {
4561 			return (cache_fplookup_dot(fpl));
4562 		}
4563 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
4564 			return (cache_fplookup_dotdot(fpl));
4565 		}
4566 	}
4567 
4568 	MPASS(!cache_fpl_isdotdot(cnp));
4569 
4570 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
4571 
4572 	CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
4573 		if (ncp->nc_dvp == dvp && ncp->nc_nlen == cnp->cn_namelen &&
4574 		    !bcmp(ncp->nc_name, cnp->cn_nameptr, ncp->nc_nlen))
4575 			break;
4576 	}
4577 
4578 	if (__predict_false(ncp == NULL)) {
4579 		return (cache_fplookup_noentry(fpl));
4580 	}
4581 
4582 	tvp = atomic_load_ptr(&ncp->nc_vp);
4583 	nc_flag = atomic_load_char(&ncp->nc_flag);
4584 	if ((nc_flag & NCF_NEGATIVE) != 0) {
4585 		return (cache_fplookup_neg(fpl, ncp, hash));
4586 	}
4587 
4588 	if (!cache_ncp_canuse(ncp)) {
4589 		return (cache_fpl_partial(fpl));
4590 	}
4591 
4592 	fpl->tvp = tvp;
4593 	fpl->tvp_seqc = vn_seqc_read_any(tvp);
4594 	if (seqc_in_modify(fpl->tvp_seqc)) {
4595 		return (cache_fpl_partial(fpl));
4596 	}
4597 
4598 	if (!cache_fplookup_vnode_supported(tvp)) {
4599 		return (cache_fpl_partial(fpl));
4600 	}
4601 
4602 	if (cache_fplookup_is_mp(fpl)) {
4603 		error = cache_fplookup_cross_mount(fpl);
4604 		if (__predict_false(error != 0)) {
4605 			return (error);
4606 		}
4607 	}
4608 
4609 	counter_u64_add(numposhits, 1);
4610 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, tvp);
4611 	return (0);
4612 }
4613 
4614 static bool
4615 cache_fplookup_mp_supported(struct mount *mp)
4616 {
4617 
4618 	MPASS(mp != NULL);
4619 	if ((mp->mnt_kern_flag & MNTK_FPLOOKUP) == 0)
4620 		return (false);
4621 	return (true);
4622 }
4623 
4624 /*
4625  * Walk up the mount stack (if any).
4626  *
4627  * Correctness is provided in the following ways:
4628  * - all vnodes are protected from freeing with SMR
4629  * - struct mount objects are type stable making them always safe to access
4630  * - stability of the particular mount is provided by busying it
4631  * - relationship between the vnode which is mounted on and the mount is
4632  *   verified with the vnode sequence counter after busying
4633  * - association between root vnode of the mount and the mount is protected
4634  *   by busy
4635  *
4636  * From that point on we can read the sequence counter of the root vnode
4637  * and get the next mount on the stack (if any) using the same protection.
4638  *
4639  * By the end of successful walk we are guaranteed the reached state was
4640  * indeed present at least at some point which matches the regular lookup.
4641  */
4642 static int __noinline
4643 cache_fplookup_climb_mount(struct cache_fpl *fpl)
4644 {
4645 	struct mount *mp, *prev_mp;
4646 	struct mount_pcpu *mpcpu, *prev_mpcpu;
4647 	struct vnode *vp;
4648 	seqc_t vp_seqc;
4649 
4650 	vp = fpl->tvp;
4651 	vp_seqc = fpl->tvp_seqc;
4652 
4653 	VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp);
4654 	mp = atomic_load_ptr(&vp->v_mountedhere);
4655 	if (__predict_false(mp == NULL)) {
4656 		return (0);
4657 	}
4658 
4659 	prev_mp = NULL;
4660 	for (;;) {
4661 		if (!vfs_op_thread_enter_crit(mp, mpcpu)) {
4662 			if (prev_mp != NULL)
4663 				vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
4664 			return (cache_fpl_partial(fpl));
4665 		}
4666 		if (prev_mp != NULL)
4667 			vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
4668 		if (!vn_seqc_consistent(vp, vp_seqc)) {
4669 			vfs_op_thread_exit_crit(mp, mpcpu);
4670 			return (cache_fpl_partial(fpl));
4671 		}
4672 		if (!cache_fplookup_mp_supported(mp)) {
4673 			vfs_op_thread_exit_crit(mp, mpcpu);
4674 			return (cache_fpl_partial(fpl));
4675 		}
4676 		vp = atomic_load_ptr(&mp->mnt_rootvnode);
4677 		if (vp == NULL || VN_IS_DOOMED(vp)) {
4678 			vfs_op_thread_exit_crit(mp, mpcpu);
4679 			return (cache_fpl_partial(fpl));
4680 		}
4681 		vp_seqc = vn_seqc_read_any(vp);
4682 		if (seqc_in_modify(vp_seqc)) {
4683 			vfs_op_thread_exit_crit(mp, mpcpu);
4684 			return (cache_fpl_partial(fpl));
4685 		}
4686 		prev_mp = mp;
4687 		prev_mpcpu = mpcpu;
4688 		mp = atomic_load_ptr(&vp->v_mountedhere);
4689 		if (mp == NULL)
4690 			break;
4691 	}
4692 
4693 	vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
4694 	fpl->tvp = vp;
4695 	fpl->tvp_seqc = vp_seqc;
4696 	return (0);
4697 }
4698 
4699 static int __noinline
4700 cache_fplookup_cross_mount(struct cache_fpl *fpl)
4701 {
4702 	struct mount *mp;
4703 	struct mount_pcpu *mpcpu;
4704 	struct vnode *vp;
4705 	seqc_t vp_seqc;
4706 
4707 	vp = fpl->tvp;
4708 	vp_seqc = fpl->tvp_seqc;
4709 
4710 	VNPASS(vp->v_type == VDIR || vp->v_type == VBAD, vp);
4711 	mp = atomic_load_ptr(&vp->v_mountedhere);
4712 	if (__predict_false(mp == NULL)) {
4713 		return (0);
4714 	}
4715 
4716 	if (!vfs_op_thread_enter_crit(mp, mpcpu)) {
4717 		return (cache_fpl_partial(fpl));
4718 	}
4719 	if (!vn_seqc_consistent(vp, vp_seqc)) {
4720 		vfs_op_thread_exit_crit(mp, mpcpu);
4721 		return (cache_fpl_partial(fpl));
4722 	}
4723 	if (!cache_fplookup_mp_supported(mp)) {
4724 		vfs_op_thread_exit_crit(mp, mpcpu);
4725 		return (cache_fpl_partial(fpl));
4726 	}
4727 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
4728 	if (__predict_false(vp == NULL || VN_IS_DOOMED(vp))) {
4729 		vfs_op_thread_exit_crit(mp, mpcpu);
4730 		return (cache_fpl_partial(fpl));
4731 	}
4732 	vp_seqc = vn_seqc_read_any(vp);
4733 	vfs_op_thread_exit_crit(mp, mpcpu);
4734 	if (seqc_in_modify(vp_seqc)) {
4735 		return (cache_fpl_partial(fpl));
4736 	}
4737 	mp = atomic_load_ptr(&vp->v_mountedhere);
4738 	if (__predict_false(mp != NULL)) {
4739 		/*
4740 		 * There are possibly more mount points on top.
4741 		 * Normally this does not happen so for simplicity just start
4742 		 * over.
4743 		 */
4744 		return (cache_fplookup_climb_mount(fpl));
4745 	}
4746 
4747 	fpl->tvp = vp;
4748 	fpl->tvp_seqc = vp_seqc;
4749 	return (0);
4750 }
4751 
4752 /*
4753  * Check if a vnode is mounted on.
4754  */
4755 static bool
4756 cache_fplookup_is_mp(struct cache_fpl *fpl)
4757 {
4758 	struct vnode *vp;
4759 
4760 	vp = fpl->tvp;
4761 	return ((vn_irflag_read(vp) & VIRF_MOUNTPOINT) != 0);
4762 }
4763 
4764 /*
4765  * Parse the path.
4766  *
4767  * The code was originally copy-pasted from regular lookup and despite
4768  * clean ups leaves performance on the table. Any modifications here
4769  * must take into account that in case off fallback the resulting
4770  * nameidata state has to be compatible with the original.
4771  */
4772 static int
4773 cache_fplookup_preparse(struct cache_fpl *fpl)
4774 {
4775 	struct nameidata *ndp;
4776 	struct componentname *cnp;
4777 
4778 	ndp = fpl->ndp;
4779 	cnp = fpl->cnp;
4780 
4781 	if (__predict_false(cnp->cn_nameptr[0] == '\0')) {
4782 		return (cache_fplookup_degenerate(fpl));
4783 	}
4784 
4785 	/*
4786 	 * By this point the shortest possible pathname is one character + nul
4787 	 * terminator, hence 2.
4788 	 */
4789 	KASSERT(ndp->ni_pathlen >= 2, ("%s: ni_pathlen %zu\n", __func__,
4790 	    ndp->ni_pathlen));
4791 
4792 	if (__predict_false(cnp->cn_nameptr[ndp->ni_pathlen - 2] == '/')) {
4793 		/*
4794 		 * TODO
4795 		 * Regular lookup performs the following:
4796 		 * *ndp->ni_next = '\0';
4797 		 * cnp->cn_flags |= TRAILINGSLASH;
4798 		 *
4799 		 * Which is problematic since it modifies data read
4800 		 * from userspace. Then if fast path lookup was to
4801 		 * abort we would have to either restore it or convey
4802 		 * the flag. Since this is a corner case just ignore
4803 		 * it for simplicity.
4804 		 */
4805 		return (cache_fpl_aborted(fpl));
4806 	}
4807 	return (0);
4808 }
4809 
4810 static int
4811 cache_fplookup_parse(struct cache_fpl *fpl)
4812 {
4813 	struct nameidata *ndp;
4814 	struct componentname *cnp;
4815 	char *cp;
4816 
4817 	ndp = fpl->ndp;
4818 	cnp = fpl->cnp;
4819 
4820 	/*
4821 	 * Find the end of this path component, it is either / or nul.
4822 	 *
4823 	 * Store / as a temporary sentinel so that we only have one character
4824 	 * to test for. Pathnames tend to be short so this should not be
4825 	 * resulting in cache misses.
4826 	 */
4827 	KASSERT(cnp->cn_nameptr[ndp->ni_pathlen - 1] == '\0',
4828 	    ("%s: expected nul at %p + %zu; string [%s]\n", __func__,
4829 	    cnp->cn_nameptr, ndp->ni_pathlen - 1, cnp->cn_nameptr));
4830 	cnp->cn_nameptr[ndp->ni_pathlen - 1] = '/';
4831 	for (cp = cnp->cn_nameptr; *cp != '/'; cp++) {
4832 		KASSERT(*cp != '\0',
4833 		    ("%s: encountered unexpected nul; string [%s]\n", __func__,
4834 		    cnp->cn_nameptr));
4835 		continue;
4836 	}
4837 	cnp->cn_nameptr[ndp->ni_pathlen - 1] = '\0';
4838 
4839 	cnp->cn_namelen = cp - cnp->cn_nameptr;
4840 	ndp->ni_pathlen -= cnp->cn_namelen;
4841 	KASSERT(ndp->ni_pathlen <= PATH_MAX,
4842 	    ("%s: ni_pathlen underflow to %zd\n", __func__, ndp->ni_pathlen));
4843 	/*
4844 	 * Hack: we have to check if the found path component's length exceeds
4845 	 * NAME_MAX. However, the condition is very rarely true and check can
4846 	 * be elided in the common case -- if an entry was found in the cache,
4847 	 * then it could not have been too long to begin with.
4848 	 */
4849 	ndp->ni_next = cp;
4850 
4851 #ifdef INVARIANTS
4852 	/*
4853 	 * Code below is only here to assure compatibility with regular lookup.
4854 	 * It covers handling of trailing slashles and names like "/", both of
4855 	 * which of can be taken care of upfront which lockless lookup does
4856 	 * in cache_fplookup_preparse. Regular lookup performs these for each
4857 	 * path component.
4858 	 */
4859 	while (*cp == '/' && (cp[1] == '/' || cp[1] == '\0')) {
4860 		cp++;
4861 		if (*cp == '\0') {
4862 			panic("%s: ran into TRAILINGSLASH handling from [%s]\n",
4863 			    __func__, cnp->cn_pnbuf);
4864 		}
4865 	}
4866 
4867 	if (cnp->cn_nameptr[0] == '\0') {
4868 		panic("%s: ran into degenerate name from [%s]\n", __func__, cnp->cn_pnbuf);
4869 	}
4870 #endif
4871 	return (0);
4872 }
4873 
4874 static void
4875 cache_fplookup_parse_advance(struct cache_fpl *fpl)
4876 {
4877 	struct nameidata *ndp;
4878 	struct componentname *cnp;
4879 
4880 	ndp = fpl->ndp;
4881 	cnp = fpl->cnp;
4882 
4883 	cnp->cn_nameptr = ndp->ni_next;
4884 	while (*cnp->cn_nameptr == '/') {
4885 		cnp->cn_nameptr++;
4886 		ndp->ni_pathlen--;
4887 	}
4888 }
4889 
4890 /*
4891  * See the API contract for VOP_FPLOOKUP_VEXEC.
4892  */
4893 static int __noinline
4894 cache_fplookup_failed_vexec(struct cache_fpl *fpl, int error)
4895 {
4896 	struct componentname *cnp;
4897 	struct vnode *dvp;
4898 	seqc_t dvp_seqc;
4899 
4900 	cnp = fpl->cnp;
4901 	dvp = fpl->dvp;
4902 	dvp_seqc = fpl->dvp_seqc;
4903 
4904 	/*
4905 	 * Hack: delayed name len checking.
4906 	 */
4907 	if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
4908 		cache_fpl_smr_exit(fpl);
4909 		return (cache_fpl_handled(fpl, ENAMETOOLONG));
4910 	}
4911 
4912 	/*
4913 	 * Hack: they may be looking up foo/bar, where foo is a
4914 	 * regular file. In such a case we need to turn ENOTDIR,
4915 	 * but we may happen to get here with a different error.
4916 	 */
4917 	if (dvp->v_type != VDIR) {
4918 		/*
4919 		 * The check here is predominantly to catch
4920 		 * EOPNOTSUPP from dead_vnodeops. If the vnode
4921 		 * gets doomed past this point it is going to
4922 		 * fail seqc verification.
4923 		 */
4924 		if (VN_IS_DOOMED(dvp)) {
4925 			return (cache_fpl_aborted(fpl));
4926 		}
4927 		error = ENOTDIR;
4928 	}
4929 
4930 	/*
4931 	 * Hack: handle O_SEARCH.
4932 	 *
4933 	 * Open Group Base Specifications Issue 7, 2018 edition states:
4934 	 * If the access mode of the open file description associated with the
4935 	 * file descriptor is not O_SEARCH, the function shall check whether
4936 	 * directory searches are permitted using the current permissions of
4937 	 * the directory underlying the file descriptor. If the access mode is
4938 	 * O_SEARCH, the function shall not perform the check.
4939 	 *
4940 	 * Regular lookup tests for the NOEXECCHECK flag for every path
4941 	 * component to decide whether to do the permission check. However,
4942 	 * since most lookups never have the flag (and when they do it is only
4943 	 * present for the first path component), lockless lookup only acts on
4944 	 * it if there is a permission problem. Here the flag is represented
4945 	 * with a boolean so that we don't have to clear it on the way out.
4946 	 *
4947 	 * For simplicity this always aborts.
4948 	 * TODO: check if this is the first lookup and ignore the permission
4949 	 * problem. Note the flag has to survive fallback (if it happens to be
4950 	 * performed).
4951 	 */
4952 	if (fpl->fsearch) {
4953 		return (cache_fpl_aborted(fpl));
4954 	}
4955 
4956 	switch (error) {
4957 	case EAGAIN:
4958 		if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4959 			error = cache_fpl_aborted(fpl);
4960 		} else {
4961 			cache_fpl_partial(fpl);
4962 		}
4963 		break;
4964 	default:
4965 		if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4966 			error = cache_fpl_aborted(fpl);
4967 		} else {
4968 			cache_fpl_smr_exit(fpl);
4969 			cache_fpl_handled(fpl, error);
4970 		}
4971 		break;
4972 	}
4973 	return (error);
4974 }
4975 
4976 static int
4977 cache_fplookup_impl(struct vnode *dvp, struct cache_fpl *fpl)
4978 {
4979 	struct nameidata *ndp;
4980 	struct componentname *cnp;
4981 	struct mount *mp;
4982 	int error;
4983 
4984 	error = CACHE_FPL_FAILED;
4985 	ndp = fpl->ndp;
4986 	cnp = fpl->cnp;
4987 
4988 	cache_fpl_checkpoint(fpl, &fpl->snd);
4989 
4990 	fpl->dvp = dvp;
4991 	fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp);
4992 	if (seqc_in_modify(fpl->dvp_seqc)) {
4993 		cache_fpl_aborted(fpl);
4994 		goto out;
4995 	}
4996 	mp = atomic_load_ptr(&dvp->v_mount);
4997 	if (__predict_false(mp == NULL || !cache_fplookup_mp_supported(mp))) {
4998 		cache_fpl_aborted(fpl);
4999 		goto out;
5000 	}
5001 
5002 	VNPASS(cache_fplookup_vnode_supported(fpl->dvp), fpl->dvp);
5003 
5004 	error = cache_fplookup_preparse(fpl);
5005 	if (__predict_false(cache_fpl_terminated(fpl))) {
5006 		goto out;
5007 	}
5008 
5009 	for (;;) {
5010 		error = cache_fplookup_parse(fpl);
5011 		if (__predict_false(error != 0)) {
5012 			break;
5013 		}
5014 
5015 		VNPASS(cache_fplookup_vnode_supported(fpl->dvp), fpl->dvp);
5016 
5017 		error = VOP_FPLOOKUP_VEXEC(fpl->dvp, cnp->cn_cred);
5018 		if (__predict_false(error != 0)) {
5019 			error = cache_fplookup_failed_vexec(fpl, error);
5020 			break;
5021 		}
5022 
5023 		error = cache_fplookup_next(fpl);
5024 		if (__predict_false(cache_fpl_terminated(fpl))) {
5025 			break;
5026 		}
5027 
5028 		VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp);
5029 
5030 		if (cache_fpl_islastcn(ndp)) {
5031 			error = cache_fplookup_final(fpl);
5032 			break;
5033 		}
5034 
5035 		if (!vn_seqc_consistent(fpl->dvp, fpl->dvp_seqc)) {
5036 			error = cache_fpl_aborted(fpl);
5037 			break;
5038 		}
5039 
5040 		fpl->dvp = fpl->tvp;
5041 		fpl->dvp_seqc = fpl->tvp_seqc;
5042 
5043 		cache_fplookup_parse_advance(fpl);
5044 		cache_fpl_checkpoint(fpl, &fpl->snd);
5045 	}
5046 out:
5047 	switch (fpl->status) {
5048 	case CACHE_FPL_STATUS_UNSET:
5049 		__assert_unreachable();
5050 		break;
5051 	case CACHE_FPL_STATUS_PARTIAL:
5052 		cache_fpl_smr_assert_entered(fpl);
5053 		return (cache_fplookup_partial_setup(fpl));
5054 	case CACHE_FPL_STATUS_ABORTED:
5055 		if (fpl->in_smr)
5056 			cache_fpl_smr_exit(fpl);
5057 		return (CACHE_FPL_FAILED);
5058 	case CACHE_FPL_STATUS_HANDLED:
5059 		MPASS(error != CACHE_FPL_FAILED);
5060 		cache_fpl_smr_assert_not_entered(fpl);
5061 		/*
5062 		 * A common error is ENOENT.
5063 		 */
5064 		if (error != 0) {
5065 			ndp->ni_dvp = NULL;
5066 			ndp->ni_vp = NULL;
5067 			cache_fpl_cleanup_cnp(cnp);
5068 			return (error);
5069 		}
5070 		ndp->ni_dvp = fpl->dvp;
5071 		ndp->ni_vp = fpl->tvp;
5072 		if (cnp->cn_flags & SAVENAME)
5073 			cnp->cn_flags |= HASBUF;
5074 		else
5075 			cache_fpl_cleanup_cnp(cnp);
5076 		return (error);
5077 	}
5078 	__assert_unreachable();
5079 }
5080 
5081 /*
5082  * Fast path lookup protected with SMR and sequence counters.
5083  *
5084  * Note: all VOP_FPLOOKUP_VEXEC routines have a comment referencing this one.
5085  *
5086  * Filesystems can opt in by setting the MNTK_FPLOOKUP flag and meeting criteria
5087  * outlined below.
5088  *
5089  * Traditional vnode lookup conceptually looks like this:
5090  *
5091  * vn_lock(current);
5092  * for (;;) {
5093  *	next = find();
5094  *	vn_lock(next);
5095  *	vn_unlock(current);
5096  *	current = next;
5097  *	if (last)
5098  *	    break;
5099  * }
5100  * return (current);
5101  *
5102  * Each jump to the next vnode is safe memory-wise and atomic with respect to
5103  * any modifications thanks to holding respective locks.
5104  *
5105  * The same guarantee can be provided with a combination of safe memory
5106  * reclamation and sequence counters instead. If all operations which affect
5107  * the relationship between the current vnode and the one we are looking for
5108  * also modify the counter, we can verify whether all the conditions held as
5109  * we made the jump. This includes things like permissions, mount points etc.
5110  * Counter modification is provided by enclosing relevant places in
5111  * vn_seqc_write_begin()/end() calls.
5112  *
5113  * Thus this translates to:
5114  *
5115  * vfs_smr_enter();
5116  * dvp_seqc = seqc_read_any(dvp);
5117  * if (seqc_in_modify(dvp_seqc)) // someone is altering the vnode
5118  *     abort();
5119  * for (;;) {
5120  * 	tvp = find();
5121  * 	tvp_seqc = seqc_read_any(tvp);
5122  * 	if (seqc_in_modify(tvp_seqc)) // someone is altering the target vnode
5123  * 	    abort();
5124  * 	if (!seqc_consistent(dvp, dvp_seqc) // someone is altering the vnode
5125  * 	    abort();
5126  * 	dvp = tvp; // we know nothing of importance has changed
5127  * 	dvp_seqc = tvp_seqc; // store the counter for the tvp iteration
5128  * 	if (last)
5129  * 	    break;
5130  * }
5131  * vget(); // secure the vnode
5132  * if (!seqc_consistent(tvp, tvp_seqc) // final check
5133  * 	    abort();
5134  * // at this point we know nothing has changed for any parent<->child pair
5135  * // as they were crossed during the lookup, meaning we matched the guarantee
5136  * // of the locked variant
5137  * return (tvp);
5138  *
5139  * The API contract for VOP_FPLOOKUP_VEXEC routines is as follows:
5140  * - they are called while within vfs_smr protection which they must never exit
5141  * - EAGAIN can be returned to denote checking could not be performed, it is
5142  *   always valid to return it
5143  * - if the sequence counter has not changed the result must be valid
5144  * - if the sequence counter has changed both false positives and false negatives
5145  *   are permitted (since the result will be rejected later)
5146  * - for simple cases of unix permission checks vaccess_vexec_smr can be used
5147  *
5148  * Caveats to watch out for:
5149  * - vnodes are passed unlocked and unreferenced with nothing stopping
5150  *   VOP_RECLAIM, in turn meaning that ->v_data can become NULL. It is advised
5151  *   to use atomic_load_ptr to fetch it.
5152  * - the aforementioned object can also get freed, meaning absent other means it
5153  *   should be protected with vfs_smr
5154  * - either safely checking permissions as they are modified or guaranteeing
5155  *   their stability is left to the routine
5156  */
5157 int
5158 cache_fplookup(struct nameidata *ndp, enum cache_fpl_status *status,
5159     struct pwd **pwdp)
5160 {
5161 	struct cache_fpl fpl;
5162 	struct pwd *pwd;
5163 	struct vnode *dvp;
5164 	struct componentname *cnp;
5165 	struct nameidata_saved orig;
5166 	int error;
5167 
5168 	MPASS(ndp->ni_lcf == 0);
5169 
5170 	fpl.status = CACHE_FPL_STATUS_UNSET;
5171 	fpl.ndp = ndp;
5172 	fpl.cnp = &ndp->ni_cnd;
5173 	MPASS(curthread == fpl.cnp->cn_thread);
5174 	KASSERT ((fpl.cnp->cn_flags & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
5175 	    ("%s: internal flags found in cn_flags %" PRIx64, __func__,
5176 	    fpl.cnp->cn_flags));
5177 
5178 	if ((fpl.cnp->cn_flags & SAVESTART) != 0)
5179 		MPASS(fpl.cnp->cn_nameiop != LOOKUP);
5180 
5181 	if (!cache_can_fplookup(&fpl)) {
5182 		SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
5183 		*status = fpl.status;
5184 		return (EOPNOTSUPP);
5185 	}
5186 
5187 	cache_fpl_checkpoint(&fpl, &orig);
5188 
5189 	cache_fpl_smr_enter_initial(&fpl);
5190 	fpl.fsearch = false;
5191 	pwd = pwd_get_smr();
5192 	fpl.pwd = pwd;
5193 	ndp->ni_rootdir = pwd->pwd_rdir;
5194 	ndp->ni_topdir = pwd->pwd_jdir;
5195 
5196 	cnp = fpl.cnp;
5197 	cnp->cn_nameptr = cnp->cn_pnbuf;
5198 	if (cnp->cn_pnbuf[0] == '/') {
5199 		dvp = cache_fpl_handle_root(&fpl);
5200 		ndp->ni_resflags |= NIRES_ABS;
5201 	} else {
5202 		if (ndp->ni_dirfd == AT_FDCWD) {
5203 			dvp = pwd->pwd_cdir;
5204 		} else {
5205 			error = cache_fplookup_dirfd(&fpl, &dvp);
5206 			if (__predict_false(error != 0)) {
5207 				goto out;
5208 			}
5209 		}
5210 	}
5211 
5212 	SDT_PROBE4(vfs, namei, lookup, entry, dvp, cnp->cn_pnbuf, cnp->cn_flags, true);
5213 
5214 	error = cache_fplookup_impl(dvp, &fpl);
5215 out:
5216 	cache_fpl_smr_assert_not_entered(&fpl);
5217 	SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
5218 
5219 	*status = fpl.status;
5220 	switch (fpl.status) {
5221 	case CACHE_FPL_STATUS_UNSET:
5222 		__assert_unreachable();
5223 		break;
5224 	case CACHE_FPL_STATUS_HANDLED:
5225 		if (error != 0)
5226 			MPASS(ndp->ni_vp == NULL);
5227 		SDT_PROBE3(vfs, namei, lookup, return, error, ndp->ni_vp, true);
5228 		break;
5229 	case CACHE_FPL_STATUS_PARTIAL:
5230 		*pwdp = fpl.pwd;
5231 		/*
5232 		 * Status restored by cache_fplookup_partial_setup.
5233 		 */
5234 		break;
5235 	case CACHE_FPL_STATUS_ABORTED:
5236 		cache_fpl_restore_abort(&fpl, &orig);
5237 		break;
5238 	}
5239 	return (error);
5240 }
5241