xref: /freebsd/sys/kern/vfs_cache.c (revision b98124e1c93789efcddaad57a8b9146ccc9a165e)
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 
35 #include "opt_ddb.h"
36 #include "opt_ktrace.h"
37 
38 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/capsicum.h>
41 #include <sys/counter.h>
42 #include <sys/filedesc.h>
43 #include <sys/fnv_hash.h>
44 #include <sys/inotify.h>
45 #include <sys/kernel.h>
46 #include <sys/ktr.h>
47 #include <sys/lock.h>
48 #include <sys/malloc.h>
49 #include <sys/fcntl.h>
50 #include <sys/jail.h>
51 #include <sys/mount.h>
52 #include <sys/namei.h>
53 #include <sys/proc.h>
54 #include <sys/seqc.h>
55 #include <sys/sdt.h>
56 #include <sys/smr.h>
57 #include <sys/smp.h>
58 #include <sys/syscallsubr.h>
59 #include <sys/sysctl.h>
60 #include <sys/sysproto.h>
61 #include <sys/vnode.h>
62 #include <ck_queue.h>
63 #ifdef KTRACE
64 #include <sys/ktrace.h>
65 #endif
66 #ifdef INVARIANTS
67 #include <machine/_inttypes.h>
68 #endif
69 
70 #include <security/audit/audit.h>
71 #include <security/mac/mac_framework.h>
72 
73 #ifdef DDB
74 #include <ddb/ddb.h>
75 #endif
76 
77 #include <vm/uma.h>
78 
79 /*
80  * High level overview of name caching in the VFS layer.
81  *
82  * Originally caching was implemented as part of UFS, later extracted to allow
83  * use by other filesystems. A decision was made to make it optional and
84  * completely detached from the rest of the kernel, which comes with limitations
85  * outlined near the end of this comment block.
86  *
87  * This fundamental choice needs to be revisited. In the meantime, the current
88  * state is described below. Significance of all notable routines is explained
89  * in comments placed above their implementation. Scattered throughout the
90  * file are TODO comments indicating shortcomings which can be fixed without
91  * reworking everything (most of the fixes will likely be reusable). Various
92  * details are omitted from this explanation to not clutter the overview, they
93  * have to be checked by reading the code and associated commentary.
94  *
95  * Keep in mind that it's individual path components which are cached, not full
96  * paths. That is, for a fully cached path "foo/bar/baz" there are 3 entries,
97  * one for each name.
98  *
99  * I. Data organization
100  *
101  * Entries are described by "struct namecache" objects and stored in a hash
102  * table. See cache_get_hash for more information.
103  *
104  * "struct vnode" contains pointers to source entries (names which can be found
105  * when traversing through said vnode), destination entries (names of that
106  * vnode (see "Limitations" for a breakdown on the subject) and a pointer to
107  * the parent vnode.
108  *
109  * The (directory vnode; name) tuple reliably determines the target entry if
110  * it exists.
111  *
112  * Since there were no small locks at the time of writing this comment (all are
113  * 32 bytes in size on LP64), the code works around the problem by introducing
114  * lock arrays to protect hash buckets and vnode lists.
115  *
116  * II. Filesystem integration
117  *
118  * Filesystems participating in name caching do the following:
119  * - set vop_lookup routine to vfs_cache_lookup
120  * - set vop_cachedlookup to a routine which can perform the lookup if the
121  *   above fails
122  * - if they support lockless lookup (see below), they set vop_fplookup_vexec
123  *   and vop_fplookup_symlink along with the MNTK_FPLOOKUP flag on the mount
124  *   point
125  * - call cache_purge or cache_vop_* routines to eliminate stale entries as
126  *   applicable
127  * - call cache_enter to add entries depending on the MAKEENTRY flag
128  *
129  * With the above in mind, there are 2 entry points when doing lookups:
130  * - ... -> namei -> cache_fplookup -- this is the default
131  * - ... -> VOP_LOOKUP -> vfs_cache_lookup -- normally only called by namei
132  *   should the above fail
133  *
134  * Example code flow how an entry is added:
135  * ... -> namei -> cache_fplookup -> cache_fplookup_noentry -> VOP_LOOKUP ->
136  * vfs_cache_lookup -> VOP_CACHEDLOOKUP -> ufs_lookup_ino -> cache_enter
137  *
138  * You may notice a degree of CPU waste in this callchain.
139  *
140  * III. Performance considerations
141  *
142  * For lockless case forward lookup avoids any writes to shared areas apart
143  * from the terminal path component. In other words non-modifying lookups of
144  * different files don't suffer any scalability problems in the namecache
145  * itself.
146  *
147  * Looking up the same file is limited by VFS and goes beyond the scope of this
148  * file.
149  *
150  * At least on amd64 the single-threaded bottleneck for long paths is hashing
151  * (see cache_get_hash). There are cases where the code issues acquire fence
152  * multiple times, they can be combined on architectures which suffer from it.
153  *
154  * For locked case each encountered vnode has to be referenced and locked in
155  * order to be handed out to the caller (normally that's namei). This
156  * introduces significant hit single-threaded and serialization multi-threaded.
157  *
158  * Reverse lookup (e.g., "getcwd") fully scales provided it is fully cached --
159  * avoids any writes to shared areas to any components.
160  *
161  * Unrelated insertions are partially serialized on updating the global entry
162  * counter and possibly serialized on colliding bucket or vnode locks.
163  *
164  * IV. Observability
165  *
166  * Several statistics are collected in the vfs.cache sysctl tree.
167  *
168  * Some of the state can be checked for with explicit dtrace probes, must of it
169  * depends on implementation details.
170  *
171  * Examples:
172  *
173  * # Check what lookups failed to be handled in a lockless manner. Column 1 is
174  * # line number, column 2 is status code (see cache_fpl_status)
175  * dtrace -n 'vfs:fplookup:lookup:done { @[arg1, arg2] = count(); }'
176  *
177  * # Histogram of lengths of names added, aggregated by which programs are doing it
178  * dtrace -n 'fbt::cache_enter_time:entry { @[execname] = quantize(args[2]->cn_namelen); }'
179  *
180  * # Same as above but only those which exceed 64 characters
181  * dtrace -n 'fbt::cache_enter_time:entry /args[2]->cn_namelen > 64/ { @[execname] = quantize(args[2]->cn_namelen); }'
182  *
183  * # Who is performing lookups with spurious slashes (e.g., "foo//bar") and what
184  * # path is it
185  * dtrace -n 'fbt::cache_fplookup_skip_slashes:entry { @[execname, stringof(args[0]->cnp->cn_pnbuf)] = count(); }'
186  *
187  * V. Limitations and implementation defects
188  *
189  * - since it is possible there is no entry for an open file, tools like
190  *   "procstat" may fail to resolve fd -> vnode -> path to anything
191  * - even if a filesystem adds an entry, it may get purged (e.g., due to memory
192  *   shortage) in which case the above problem applies
193  * - hardlinks are not tracked, thus if a vnode is reachable in more than one
194  *   way, resolving a name may return a different path than the one used to
195  *   open it (even if said path is still valid)
196  * - by default entries are not added for newly created files
197  * - adding an entry may need to evict negative entry first, which happens in 2
198  *   distinct places (evicting on lookup, adding in a later VOP) making it
199  *   impossible to simply reuse it
200  * - there is a simple scheme to evict negative entries as the cache is approaching
201  *   its capacity, but it is very unclear if doing so is a good idea to begin with
202  * - vnodes are subject to being recycled even if target inode is left in memory,
203  *   which loses the name cache entries when it perhaps should not. in case of tmpfs
204  *   names get duplicated -- kept by filesystem itself and namecache separately
205  * - vnode reclamation (see vnlru in kern/vfs_subr.c) defaults to skipping
206  *   directories for this very reason, which arguably further reducing quality
207  *   of vnode LRU. Per the above this is done to avoid breaking vnode -> path
208  *   resolution (it becomes expensive for directories and impossible for the rest)
209  *   This would not be a factor if namecache entries could persist without vnodes.
210  * - struct namecache has a fixed size and comes in 2 variants, often wasting
211  *   space.  now hard to replace with malloc due to dependence on SMR, which
212  *   requires UMA zones to opt in
213  * - lack of better integration with the kernel also turns nullfs into a layered
214  *   filesystem instead of something which can take advantage of caching
215  *
216  * Appendix A: where is the time lost, expanding on paragraph III
217  *
218  * While some care went into optimizing lookups, there is still plenty of
219  * performance left on the table, most notably from single-threaded standpoint.
220  * Below is a woefully incomplete list of changes which can help.  Ideas are
221  * mostly sketched out, no claim is made all kinks or prerequisites are laid
222  * out. The name of the game is eliding branches altogether and hopefully some
223  * of memory accesses.
224  *
225  * Note there is performance lost all over VFS.
226  *
227  * === SMR-only lookup
228  *
229  * For commonly used ops like stat(2), when the terminal vnode *is* cached,
230  * lockless lookup could refrain from refing/locking the found vnode and
231  * instead return while within the SMR section. Then a call to, say,
232  * vop_stat_smr could do the work (or fail with EAGAIN), finally the result
233  * would be validated with seqc not changing. This would be faster
234  * single-threaded as it dodges atomics and would provide full scalability for
235  * multicore uses. This would *not* work for open(2) or other calls which need
236  * the vnode to hang around for the long haul, but would work for aforementioned
237  * stat(2) but also access(2), readlink(2), realpathat(2) and probably more.
238  *
239  * === copyinstr
240  *
241  * On all architectures it operates one byte at a time, while it could be
242  * word-sized instead thanks to the Mycroft trick.
243  *
244  * API itself is rather pessimal for path lookup, accepting arbitrary sizes and
245  * *optionally* filling in the length parameter.
246  *
247  * Instead a new routine (copyinpath?) could be introduced, demanding a buffer
248  * size which is a multiply of the word (and never zero), with the length
249  * always returned. On top of it the routine could be allowed to transform the
250  * buffer in arbitrary ways, most notably writing past the found length (not to
251  * be confused with writing past buffer size) -- this would allow word-sized
252  * movs while checking for '\0' later.
253  *
254  * === detour through namei
255  *
256  * Currently one suffers being called from namei, which then has to check if
257  * things worked out locklessly. Instead the lockless lookup could be the
258  * actual entry point which calls what is currently namei as a fallback.
259  *
260  * It could be hotpatched if lockless lookup is disabled.
261  *
262  * === avoidable branches in cache_can_fplookup
263  *
264  * The cache_fast_lookup_enabled flag check could be hotpatchable (in fact if
265  * this is off, none of fplookup code should execute, see above).
266  *
267  * Both audit and capsicum branches can be combined into one, but it requires
268  * paying off a lot of tech debt first.
269  *
270  * ni_startdir could be indicated with a flag in cn_flags, eliminating the
271  * branch.
272  *
273  * === mount stacks
274  *
275  * Crossing a mount requires checking if perhaps something is mounted on top.
276  * Instead, an additional entry could be added to struct mount with a pointer
277  * to the final mount on the stack. This would be recalculated on each
278  * mount/unmount.
279  *
280  * === root vnodes
281  *
282  * It could become part of the API contract to *always* have a rootvnode set in
283  * mnt_rootvnode. Such vnodes are annotated with VV_ROOT and vnlru would have
284  * to be modified to always skip them.
285  *
286  * === inactive on v_usecount reaching 0
287  *
288  * VOP_NEED_INACTIVE should not exist. Filesystems can indicate need for such
289  * processing with a bit in usecount and adding a hold count. Then vput fast path
290  * would become as simple as (ACHTUNG: locking ignored):
291  *
292  * ref = atomic_fetchadd_int(&vp->v_count, -1) - 1;
293  * if ((ref & MAGIC_BIT) == 0) // common case
294  *	return;
295  * if (ref != 0) // the bit is set but this was not the last user
296  *	return;
297  * // do inactive here
298  *
299  * Also see below.
300  *
301  * === v_holdcnt
302  *
303  * Hold count should probably get eliminated, but one can argue it is a useful
304  * feature. Even if so, handling of v_usecount could be decoupled from it --
305  * vnlru et al would consider the vnode not-freeable if has either hold or
306  * usecount on it.
307  *
308  * This would eliminate 2 atomics in the common case of securing a vnode and
309  * undoing it.
310  */
311 
312 static SYSCTL_NODE(_vfs, OID_AUTO, cache, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
313     "Name cache");
314 
315 SDT_PROVIDER_DECLARE(vfs);
316 SDT_PROBE_DEFINE3(vfs, namecache, enter, done, "struct vnode *", "char *",
317     "struct vnode *");
318 SDT_PROBE_DEFINE3(vfs, namecache, enter, duplicate, "struct vnode *", "char *",
319     "struct vnode *");
320 SDT_PROBE_DEFINE2(vfs, namecache, enter_negative, done, "struct vnode *",
321     "char *");
322 SDT_PROBE_DEFINE2(vfs, namecache, fullpath_smr, hit, "struct vnode *",
323     "const char *");
324 SDT_PROBE_DEFINE4(vfs, namecache, fullpath_smr, miss, "struct vnode *",
325     "struct namecache *", "int", "int");
326 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, entry, "struct vnode *");
327 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, hit, "struct vnode *",
328     "char *", "struct vnode *");
329 SDT_PROBE_DEFINE1(vfs, namecache, fullpath, miss, "struct vnode *");
330 SDT_PROBE_DEFINE3(vfs, namecache, fullpath, return, "int",
331     "struct vnode *", "char *");
332 SDT_PROBE_DEFINE3(vfs, namecache, lookup, hit, "struct vnode *", "char *",
333     "struct vnode *");
334 SDT_PROBE_DEFINE2(vfs, namecache, lookup, hit__negative,
335     "struct vnode *", "char *");
336 SDT_PROBE_DEFINE2(vfs, namecache, lookup, miss, "struct vnode *",
337     "char *");
338 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, hit, "struct vnode *",
339     "struct componentname *");
340 SDT_PROBE_DEFINE2(vfs, namecache, removecnp, miss, "struct vnode *",
341     "struct componentname *");
342 SDT_PROBE_DEFINE3(vfs, namecache, purge, done, "struct vnode *", "size_t", "size_t");
343 SDT_PROBE_DEFINE1(vfs, namecache, purge, batch, "int");
344 SDT_PROBE_DEFINE1(vfs, namecache, purge_negative, done, "struct vnode *");
345 SDT_PROBE_DEFINE1(vfs, namecache, purgevfs, done, "struct mount *");
346 SDT_PROBE_DEFINE3(vfs, namecache, zap, done, "struct vnode *", "char *",
347     "struct vnode *");
348 SDT_PROBE_DEFINE2(vfs, namecache, zap_negative, done, "struct vnode *",
349     "char *");
350 SDT_PROBE_DEFINE2(vfs, namecache, evict_negative, done, "struct vnode *",
351     "char *");
352 SDT_PROBE_DEFINE1(vfs, namecache, symlink, alloc__fail, "size_t");
353 
354 SDT_PROBE_DEFINE3(vfs, fplookup, lookup, done, "struct nameidata *", "int",
355     "enum cache_fpl_status");
356 SDT_PROBE_DECLARE(vfs, namei, lookup, entry);
357 SDT_PROBE_DECLARE(vfs, namei, lookup, return);
358 
359 static char __read_frequently cache_fast_lookup_enabled = true;
360 
361 /*
362  * This structure describes the elements in the cache of recent
363  * names looked up by namei.
364  */
365 struct negstate {
366 	u_char neg_flag;
367 	u_char neg_hit;
368 };
369 _Static_assert(sizeof(struct negstate) <= sizeof(struct vnode *),
370     "the state must fit in a union with a pointer without growing it");
371 
372 struct	namecache {
373 	LIST_ENTRY(namecache) nc_src;	/* source vnode list */
374 	TAILQ_ENTRY(namecache) nc_dst;	/* destination vnode list */
375 	CK_SLIST_ENTRY(namecache) nc_hash;/* hash chain */
376 	struct	vnode *nc_dvp;		/* vnode of parent of name */
377 	union {
378 		struct	vnode *nu_vp;	/* vnode the name refers to */
379 		struct	negstate nu_neg;/* negative entry state */
380 	} n_un;
381 	u_char	nc_flag;		/* flag bits */
382 	u_char	nc_nlen;		/* length of name */
383 	char	nc_name[];		/* segment name + nul */
384 };
385 
386 /*
387  * struct namecache_ts repeats struct namecache layout up to the
388  * nc_nlen member.
389  * struct namecache_ts is used in place of struct namecache when time(s) need
390  * to be stored.  The nc_dotdottime field is used when a cache entry is mapping
391  * both a non-dotdot directory name plus dotdot for the directory's
392  * parent.
393  *
394  * See below for alignment requirement.
395  */
396 struct	namecache_ts {
397 	struct	timespec nc_time;	/* timespec provided by fs */
398 	struct	timespec nc_dotdottime;	/* dotdot timespec provided by fs */
399 	int	nc_ticks;		/* ticks value when entry was added */
400 	int	nc_pad;
401 	struct namecache nc_nc;
402 };
403 
404 TAILQ_HEAD(cache_freebatch, namecache);
405 
406 /*
407  * At least mips n32 performs 64-bit accesses to timespec as found
408  * in namecache_ts and requires them to be aligned. Since others
409  * may be in the same spot suffer a little bit and enforce the
410  * alignment for everyone. Note this is a nop for 64-bit platforms.
411  */
412 #define CACHE_ZONE_ALIGNMENT	UMA_ALIGNOF(time_t)
413 
414 /*
415  * TODO: the initial value of CACHE_PATH_CUTOFF was inherited from the
416  * 4.4 BSD codebase. Later on struct namecache was tweaked to become
417  * smaller and the value was bumped to retain the total size, but it
418  * was never re-evaluated for suitability. A simple test counting
419  * lengths during package building shows that the value of 45 covers
420  * about 86% of all added entries, reaching 99% at 65.
421  *
422  * Regardless of the above, use of dedicated zones instead of malloc may be
423  * inducing additional waste. This may be hard to address as said zones are
424  * tied to VFS SMR. Even if retaining them, the current split should be
425  * re-evaluated.
426  */
427 #ifdef __LP64__
428 #define	CACHE_PATH_CUTOFF	45
429 #define	CACHE_LARGE_PAD		6
430 #else
431 #define	CACHE_PATH_CUTOFF	41
432 #define	CACHE_LARGE_PAD		2
433 #endif
434 
435 #define CACHE_ZONE_SMALL_SIZE		(offsetof(struct namecache, nc_name) + CACHE_PATH_CUTOFF + 1)
436 #define CACHE_ZONE_SMALL_TS_SIZE	(offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_SMALL_SIZE)
437 #define CACHE_ZONE_LARGE_SIZE		(offsetof(struct namecache, nc_name) + NAME_MAX + 1 + CACHE_LARGE_PAD)
438 #define CACHE_ZONE_LARGE_TS_SIZE	(offsetof(struct namecache_ts, nc_nc) + CACHE_ZONE_LARGE_SIZE)
439 
440 _Static_assert((CACHE_ZONE_SMALL_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
441 _Static_assert((CACHE_ZONE_SMALL_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
442 _Static_assert((CACHE_ZONE_LARGE_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
443 _Static_assert((CACHE_ZONE_LARGE_TS_SIZE % (CACHE_ZONE_ALIGNMENT + 1)) == 0, "bad zone size");
444 
445 #define	nc_vp		n_un.nu_vp
446 #define	nc_neg		n_un.nu_neg
447 
448 /*
449  * Flags in namecache.nc_flag
450  */
451 #define NCF_WHITE	0x01
452 #define NCF_ISDOTDOT	0x02
453 #define	NCF_TS		0x04
454 #define	NCF_DTS		0x08
455 #define	NCF_DVDROP	0x10
456 #define	NCF_NEGATIVE	0x20
457 #define	NCF_INVALID	0x40
458 #define	NCF_WIP		0x80
459 
460 /*
461  * Flags in negstate.neg_flag
462  */
463 #define NEG_HOT		0x01
464 
465 static bool	cache_neg_evict_cond(u_long lnumcache);
466 
467 /*
468  * Mark an entry as invalid.
469  *
470  * This is called before it starts getting deconstructed.
471  */
472 static void
cache_ncp_invalidate(struct namecache * ncp)473 cache_ncp_invalidate(struct namecache *ncp)
474 {
475 
476 	KASSERT((ncp->nc_flag & NCF_INVALID) == 0,
477 	    ("%s: entry %p already invalid", __func__, ncp));
478 	atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_INVALID);
479 	atomic_thread_fence_rel();
480 }
481 
482 /*
483  * Does this entry match the given directory and name?
484  */
485 static bool
cache_ncp_match(struct namecache * ncp,struct vnode * dvp,struct componentname * cnp)486 cache_ncp_match(struct namecache *ncp, struct vnode *dvp,
487     struct componentname *cnp)
488 {
489 	return (ncp->nc_dvp == dvp &&
490 	    ncp->nc_nlen == cnp->cn_namelen &&
491 	    bcmp(ncp->nc_name, cnp->cn_nameptr, cnp->cn_namelen) == 0);
492 }
493 
494 /*
495  * Check whether the entry can be safely used.
496  *
497  * All places which elide locks are supposed to call this after they are
498  * done with reading from an entry.
499  */
500 #define cache_ncp_canuse(ncp)	({					\
501 	struct namecache *_ncp = (ncp);					\
502 	u_char _nc_flag;						\
503 									\
504 	atomic_thread_fence_acq();					\
505 	_nc_flag = atomic_load_char(&_ncp->nc_flag);			\
506 	__predict_true((_nc_flag & (NCF_INVALID | NCF_WIP)) == 0);	\
507 })
508 
509 /*
510  * Like the above but also checks NCF_WHITE.
511  */
512 #define cache_fpl_neg_ncp_canuse(ncp)	({				\
513 	struct namecache *_ncp = (ncp);					\
514 	u_char _nc_flag;						\
515 									\
516 	atomic_thread_fence_acq();					\
517 	_nc_flag = atomic_load_char(&_ncp->nc_flag);			\
518 	__predict_true((_nc_flag & (NCF_INVALID | NCF_WIP | NCF_WHITE)) == 0);	\
519 })
520 
521 VFS_SMR_DECLARE;
522 
523 static SYSCTL_NODE(_vfs_cache, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
524     "Name cache parameters");
525 
526 static u_int __read_mostly	ncsize; /* the size as computed on creation or resizing */
527 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, size, CTLFLAG_RD, &ncsize, 0,
528     "Total namecache capacity");
529 
530 u_int ncsizefactor = 2;
531 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, sizefactor, CTLFLAG_RW, &ncsizefactor, 0,
532     "Size factor for namecache");
533 
534 static u_long __read_mostly	ncnegfactor = 5; /* ratio of negative entries */
535 SYSCTL_ULONG(_vfs_cache_param, OID_AUTO, negfactor, CTLFLAG_RW, &ncnegfactor, 0,
536     "Ratio of negative namecache entries");
537 
538 /*
539  * Negative entry % of namecache capacity above which automatic eviction is allowed.
540  *
541  * Check cache_neg_evict_cond for details.
542  */
543 static u_int ncnegminpct = 3;
544 
545 static u_int __read_mostly     neg_min; /* the above recomputed against ncsize */
546 SYSCTL_UINT(_vfs_cache_param, OID_AUTO, negmin, CTLFLAG_RD, &neg_min, 0,
547     "Negative entry count above which automatic eviction is allowed");
548 
549 /*
550  * Structures associated with name caching.
551  */
552 #define NCHHASH(hash) \
553 	(&nchashtbl[(hash) & nchash])
554 static __read_mostly CK_SLIST_HEAD(nchashhead, namecache) *nchashtbl;/* Hash Table */
555 static u_long __read_mostly	nchash;			/* size of hash table */
556 SYSCTL_ULONG(_debug, OID_AUTO, nchash, CTLFLAG_RD, &nchash, 0,
557     "Size of namecache hash table");
558 static u_long __exclusive_cache_line	numneg;	/* number of negative entries allocated */
559 static u_long __exclusive_cache_line	numcache;/* number of cache entries allocated */
560 
561 struct nchstats	nchstats;		/* cache effectiveness statistics */
562 
563 static u_int __exclusive_cache_line neg_cycle;
564 
565 #define ncneghash	3
566 #define	numneglists	(ncneghash + 1)
567 
568 struct neglist {
569 	struct mtx		nl_evict_lock;
570 	struct mtx		nl_lock __aligned(CACHE_LINE_SIZE);
571 	TAILQ_HEAD(, namecache) nl_list;
572 	TAILQ_HEAD(, namecache) nl_hotlist;
573 	u_long			nl_hotnum;
574 } __aligned(CACHE_LINE_SIZE);
575 
576 static struct neglist neglists[numneglists];
577 
578 static inline struct neglist *
NCP2NEGLIST(struct namecache * ncp)579 NCP2NEGLIST(struct namecache *ncp)
580 {
581 
582 	return (&neglists[(((uintptr_t)(ncp) >> 8) & ncneghash)]);
583 }
584 
585 static inline struct negstate *
NCP2NEGSTATE(struct namecache * ncp)586 NCP2NEGSTATE(struct namecache *ncp)
587 {
588 
589 	MPASS(atomic_load_char(&ncp->nc_flag) & NCF_NEGATIVE);
590 	return (&ncp->nc_neg);
591 }
592 
593 #define	numbucketlocks (ncbuckethash + 1)
594 static u_int __read_mostly  ncbuckethash;
595 static struct mtx_padalign __read_mostly  *bucketlocks;
596 #define	HASH2BUCKETLOCK(hash) \
597 	((struct mtx *)(&bucketlocks[((hash) & ncbuckethash)]))
598 
599 #define	numvnodelocks (ncvnodehash + 1)
600 static u_int __read_mostly  ncvnodehash;
601 static struct mtx __read_mostly *vnodelocks;
602 static inline struct mtx *
VP2VNODELOCK(struct vnode * vp)603 VP2VNODELOCK(struct vnode *vp)
604 {
605 
606 	return (&vnodelocks[(((uintptr_t)(vp) >> 8) & ncvnodehash)]);
607 }
608 
609 /*
610  * Search the hash table for a namecache entry.  Either the corresponding bucket
611  * must be locked, or the caller must be in an SMR read section.
612  */
613 static struct namecache *
cache_ncp_find(struct vnode * dvp,struct componentname * cnp,uint32_t hash)614 cache_ncp_find(struct vnode *dvp, struct componentname *cnp, uint32_t hash)
615 {
616 	struct namecache *ncp;
617 
618 	KASSERT(mtx_owned(HASH2BUCKETLOCK(hash)) || VFS_SMR_ENTERED(),
619 	    ("%s: hash %u not locked", __func__, hash));
620 	CK_SLIST_FOREACH(ncp, NCHHASH(hash), nc_hash) {
621 		if (cache_ncp_match(ncp, dvp, cnp))
622 			break;
623 	}
624 	return (ncp);
625 }
626 
627 static void
cache_out_ts(struct namecache * ncp,struct timespec * tsp,int * ticksp)628 cache_out_ts(struct namecache *ncp, struct timespec *tsp, int *ticksp)
629 {
630 	struct namecache_ts *ncp_ts;
631 
632 	KASSERT((ncp->nc_flag & NCF_TS) != 0 ||
633 	    (tsp == NULL && ticksp == NULL),
634 	    ("No NCF_TS"));
635 
636 	if (tsp == NULL)
637 		return;
638 
639 	ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
640 	*tsp = ncp_ts->nc_time;
641 	*ticksp = ncp_ts->nc_ticks;
642 }
643 
644 #ifdef DEBUG_CACHE
645 static int __read_mostly	doingcache = 1;	/* 1 => enable the cache */
646 SYSCTL_INT(_debug, OID_AUTO, vfscache, CTLFLAG_RW, &doingcache, 0,
647     "VFS namecache enabled");
648 #endif
649 
650 /* Export size information to userland */
651 SYSCTL_SIZEOF_STRUCT(namecache);
652 
653 /*
654  * The new name cache statistics
655  */
656 static SYSCTL_NODE(_vfs_cache, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
657     "Name cache statistics");
658 
659 #define STATNODE_ULONG(name, varname, descr)					\
660 	SYSCTL_ULONG(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
661 #define STATNODE_COUNTER(name, varname, descr)					\
662 	static COUNTER_U64_DEFINE_EARLY(varname);				\
663 	SYSCTL_COUNTER_U64(_vfs_cache_stats, OID_AUTO, name, CTLFLAG_RD, &varname, \
664 	    descr);
665 STATNODE_ULONG(neg, numneg, "Number of negative cache entries");
666 STATNODE_ULONG(count, numcache, "Number of cache entries");
667 STATNODE_COUNTER(heldvnodes, numcachehv, "Number of namecache entries with vnodes held");
668 STATNODE_COUNTER(drops, numdrops, "Number of dropped entries due to reaching the limit");
669 STATNODE_COUNTER(miss, nummiss, "Number of cache misses");
670 STATNODE_COUNTER(misszap, nummisszap, "Number of cache misses we do not want to cache");
671 STATNODE_COUNTER(poszaps, numposzaps,
672     "Number of cache hits (positive) we do not want to cache");
673 STATNODE_COUNTER(poshits, numposhits, "Number of cache hits (positive)");
674 STATNODE_COUNTER(negzaps, numnegzaps,
675     "Number of cache hits (negative) we do not want to cache");
676 STATNODE_COUNTER(neghits, numneghits, "Number of cache hits (negative)");
677 /* These count for vn_getcwd(), too. */
678 STATNODE_COUNTER(fullpathcalls, numfullpathcalls, "Number of fullpath search calls");
679 STATNODE_COUNTER(fullpathfail2, numfullpathfail2,
680     "Number of fullpath search errors (VOP_VPTOCNP failures)");
681 STATNODE_COUNTER(fullpathfail4, numfullpathfail4, "Number of fullpath search errors (ENOMEM)");
682 STATNODE_COUNTER(fullpathfound, numfullpathfound, "Number of successful fullpath calls");
683 STATNODE_COUNTER(symlinktoobig, symlinktoobig, "Number of times symlink did not fit the cache");
684 
685 /*
686  * Debug or developer statistics.
687  */
688 static SYSCTL_NODE(_vfs_cache, OID_AUTO, debug, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
689     "Name cache debugging");
690 #define DEBUGNODE_ULONG(name, varname, descr)					\
691 	SYSCTL_ULONG(_vfs_cache_debug, OID_AUTO, name, CTLFLAG_RD, &varname, 0, descr);
692 static u_long zap_bucket_relock_success;
693 DEBUGNODE_ULONG(zap_bucket_relock_success, zap_bucket_relock_success,
694     "Number of successful removals after relocking");
695 static u_long zap_bucket_fail;
696 DEBUGNODE_ULONG(zap_bucket_fail, zap_bucket_fail, "");
697 static u_long zap_bucket_fail2;
698 DEBUGNODE_ULONG(zap_bucket_fail2, zap_bucket_fail2, "");
699 static u_long cache_lock_vnodes_cel_3_failures;
700 DEBUGNODE_ULONG(vnodes_cel_3_failures, cache_lock_vnodes_cel_3_failures,
701     "Number of times 3-way vnode locking failed");
702 
703 static void cache_zap_locked(struct namecache *ncp);
704 static int vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
705     char **retbuf, size_t *buflen, size_t addend);
706 static int vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf,
707     char **retbuf, size_t *buflen);
708 static int vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf,
709     char **retbuf, size_t *len, size_t addend);
710 
711 static MALLOC_DEFINE(M_VFSCACHE, "vfscache", "VFS name cache entries");
712 
713 static inline void
cache_assert_vlp_locked(struct mtx * vlp)714 cache_assert_vlp_locked(struct mtx *vlp)
715 {
716 
717 	if (vlp != NULL)
718 		mtx_assert(vlp, MA_OWNED);
719 }
720 
721 static inline void
cache_assert_vnode_locked(struct vnode * vp)722 cache_assert_vnode_locked(struct vnode *vp)
723 {
724 	struct mtx *vlp;
725 
726 	vlp = VP2VNODELOCK(vp);
727 	cache_assert_vlp_locked(vlp);
728 }
729 
730 /*
731  * Directory vnodes with entries are held for two reasons:
732  * 1. make them less of a target for reclamation in vnlru
733  * 2. suffer smaller performance penalty in locked lookup as requeieing is avoided
734  *
735  * It will be feasible to stop doing it altogether if all filesystems start
736  * supporting lockless lookup.
737  */
738 static void
cache_hold_vnode(struct vnode * vp)739 cache_hold_vnode(struct vnode *vp)
740 {
741 
742 	cache_assert_vnode_locked(vp);
743 	VNPASS(LIST_EMPTY(&vp->v_cache_src), vp);
744 	vhold(vp);
745 	counter_u64_add(numcachehv, 1);
746 }
747 
748 static void
cache_drop_vnode(struct vnode * vp)749 cache_drop_vnode(struct vnode *vp)
750 {
751 
752 	/*
753 	 * Called after all locks are dropped, meaning we can't assert
754 	 * on the state of v_cache_src.
755 	 */
756 	vdrop(vp);
757 	counter_u64_add(numcachehv, -1);
758 }
759 
760 /*
761  * UMA zones.
762  */
763 static uma_zone_t __read_mostly cache_zone_small;
764 static uma_zone_t __read_mostly cache_zone_small_ts;
765 static uma_zone_t __read_mostly cache_zone_large;
766 static uma_zone_t __read_mostly cache_zone_large_ts;
767 
768 char *
cache_symlink_alloc(size_t size,int flags)769 cache_symlink_alloc(size_t size, int flags)
770 {
771 
772 	if (size < CACHE_ZONE_SMALL_SIZE) {
773 		return (uma_zalloc_smr(cache_zone_small, flags));
774 	}
775 	if (size < CACHE_ZONE_LARGE_SIZE) {
776 		return (uma_zalloc_smr(cache_zone_large, flags));
777 	}
778 	counter_u64_add(symlinktoobig, 1);
779 	SDT_PROBE1(vfs, namecache, symlink, alloc__fail, size);
780 	return (NULL);
781 }
782 
783 void
cache_symlink_free(char * string,size_t size)784 cache_symlink_free(char *string, size_t size)
785 {
786 
787 	MPASS(string != NULL);
788 	KASSERT(size < CACHE_ZONE_LARGE_SIZE,
789 	    ("%s: size %zu too big", __func__, size));
790 
791 	if (size < CACHE_ZONE_SMALL_SIZE) {
792 		uma_zfree_smr(cache_zone_small, string);
793 		return;
794 	}
795 	if (size < CACHE_ZONE_LARGE_SIZE) {
796 		uma_zfree_smr(cache_zone_large, string);
797 		return;
798 	}
799 	__assert_unreachable();
800 }
801 
802 static struct namecache *
cache_alloc_uma(int len,bool ts)803 cache_alloc_uma(int len, bool ts)
804 {
805 	struct namecache_ts *ncp_ts;
806 	struct namecache *ncp;
807 
808 	if (__predict_false(ts)) {
809 		if (len <= CACHE_PATH_CUTOFF)
810 			ncp_ts = uma_zalloc_smr(cache_zone_small_ts, M_WAITOK);
811 		else
812 			ncp_ts = uma_zalloc_smr(cache_zone_large_ts, M_WAITOK);
813 		ncp = &ncp_ts->nc_nc;
814 	} else {
815 		if (len <= CACHE_PATH_CUTOFF)
816 			ncp = uma_zalloc_smr(cache_zone_small, M_WAITOK);
817 		else
818 			ncp = uma_zalloc_smr(cache_zone_large, M_WAITOK);
819 	}
820 	return (ncp);
821 }
822 
823 static void
cache_free_uma(struct namecache * ncp)824 cache_free_uma(struct namecache *ncp)
825 {
826 	struct namecache_ts *ncp_ts;
827 
828 	if (__predict_false(ncp->nc_flag & NCF_TS)) {
829 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
830 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
831 			uma_zfree_smr(cache_zone_small_ts, ncp_ts);
832 		else
833 			uma_zfree_smr(cache_zone_large_ts, ncp_ts);
834 	} else {
835 		if (ncp->nc_nlen <= CACHE_PATH_CUTOFF)
836 			uma_zfree_smr(cache_zone_small, ncp);
837 		else
838 			uma_zfree_smr(cache_zone_large, ncp);
839 	}
840 }
841 
842 static struct namecache *
cache_alloc(int len,bool ts)843 cache_alloc(int len, bool ts)
844 {
845 	u_long lnumcache;
846 
847 	/*
848 	 * Avoid blowout in namecache entries.
849 	 *
850 	 * Bugs:
851 	 * 1. filesystems may end up trying to add an already existing entry
852 	 * (for example this can happen after a cache miss during concurrent
853 	 * lookup), in which case we will call cache_neg_evict despite not
854 	 * adding anything.
855 	 * 2. the routine may fail to free anything and no provisions are made
856 	 * to make it try harder (see the inside for failure modes)
857 	 * 3. it only ever looks at negative entries.
858 	 */
859 	lnumcache = atomic_fetchadd_long(&numcache, 1) + 1;
860 	if (cache_neg_evict_cond(lnumcache)) {
861 		lnumcache = atomic_load_long(&numcache);
862 	}
863 	if (__predict_false(lnumcache >= ncsize)) {
864 		atomic_subtract_long(&numcache, 1);
865 		counter_u64_add(numdrops, 1);
866 		return (NULL);
867 	}
868 	return (cache_alloc_uma(len, ts));
869 }
870 
871 static void
cache_free(struct namecache * ncp)872 cache_free(struct namecache *ncp)
873 {
874 
875 	MPASS(ncp != NULL);
876 	if ((ncp->nc_flag & NCF_DVDROP) != 0) {
877 		cache_drop_vnode(ncp->nc_dvp);
878 	}
879 	cache_free_uma(ncp);
880 	atomic_subtract_long(&numcache, 1);
881 }
882 
883 static void
cache_free_batch(struct cache_freebatch * batch)884 cache_free_batch(struct cache_freebatch *batch)
885 {
886 	struct namecache *ncp, *nnp;
887 	int i;
888 
889 	i = 0;
890 	if (TAILQ_EMPTY(batch))
891 		goto out;
892 	TAILQ_FOREACH_SAFE(ncp, batch, nc_dst, nnp) {
893 		if ((ncp->nc_flag & NCF_DVDROP) != 0) {
894 			cache_drop_vnode(ncp->nc_dvp);
895 		}
896 		cache_free_uma(ncp);
897 		i++;
898 	}
899 	atomic_subtract_long(&numcache, i);
900 out:
901 	SDT_PROBE1(vfs, namecache, purge, batch, i);
902 }
903 
904 /*
905  * Hashing.
906  *
907  * The code was made to use FNV in 2001 and this choice needs to be revisited.
908  *
909  * Short summary of the difficulty:
910  * The longest name which can be inserted is NAME_MAX characters in length (or
911  * 255 at the time of writing this comment), while majority of names used in
912  * practice are significantly shorter (mostly below 10). More importantly
913  * majority of lookups performed find names are even shorter than that.
914  *
915  * This poses a problem where hashes which do better than FNV past word size
916  * (or so) tend to come with additional overhead when finalizing the result,
917  * making them noticeably slower for the most commonly used range.
918  *
919  * Consider a path like: /usr/obj/usr/src/sys/amd64/GENERIC/vnode_if.c
920  *
921  * When looking it up the most time consuming part by a large margin (at least
922  * on amd64) is hashing.  Replacing FNV with something which pessimizes short
923  * input would make the slowest part stand out even more.
924  */
925 
926 /*
927  * TODO: With the value stored we can do better than computing the hash based
928  * on the address.
929  */
930 static void
cache_prehash(struct vnode * vp)931 cache_prehash(struct vnode *vp)
932 {
933 
934 	vp->v_nchash = fnv_32_buf(&vp, sizeof(vp), FNV1_32_INIT);
935 }
936 
937 static uint32_t
cache_get_hash(char * name,u_char len,struct vnode * dvp)938 cache_get_hash(char *name, u_char len, struct vnode *dvp)
939 {
940 
941 	return (fnv_32_buf(name, len, dvp->v_nchash));
942 }
943 
944 static uint32_t
cache_get_hash_iter_start(struct vnode * dvp)945 cache_get_hash_iter_start(struct vnode *dvp)
946 {
947 
948 	return (dvp->v_nchash);
949 }
950 
951 static uint32_t
cache_get_hash_iter(char c,uint32_t hash)952 cache_get_hash_iter(char c, uint32_t hash)
953 {
954 
955 	return (fnv_32_buf(&c, 1, hash));
956 }
957 
958 static uint32_t
cache_get_hash_iter_finish(uint32_t hash)959 cache_get_hash_iter_finish(uint32_t hash)
960 {
961 
962 	return (hash);
963 }
964 
965 static inline struct nchashhead *
NCP2BUCKET(struct namecache * ncp)966 NCP2BUCKET(struct namecache *ncp)
967 {
968 	uint32_t hash;
969 
970 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
971 	return (NCHHASH(hash));
972 }
973 
974 static inline struct mtx *
NCP2BUCKETLOCK(struct namecache * ncp)975 NCP2BUCKETLOCK(struct namecache *ncp)
976 {
977 	uint32_t hash;
978 
979 	hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen, ncp->nc_dvp);
980 	return (HASH2BUCKETLOCK(hash));
981 }
982 
983 #ifdef INVARIANTS
984 static void
cache_assert_bucket_locked(struct namecache * ncp)985 cache_assert_bucket_locked(struct namecache *ncp)
986 {
987 	struct mtx *blp;
988 
989 	blp = NCP2BUCKETLOCK(ncp);
990 	mtx_assert(blp, MA_OWNED);
991 }
992 
993 static void
cache_assert_bucket_unlocked(struct namecache * ncp)994 cache_assert_bucket_unlocked(struct namecache *ncp)
995 {
996 	struct mtx *blp;
997 
998 	blp = NCP2BUCKETLOCK(ncp);
999 	mtx_assert(blp, MA_NOTOWNED);
1000 }
1001 #else
1002 #define cache_assert_bucket_locked(x) do { } while (0)
1003 #define cache_assert_bucket_unlocked(x) do { } while (0)
1004 #endif
1005 
1006 #define cache_sort_vnodes(x, y)	_cache_sort_vnodes((void **)(x), (void **)(y))
1007 static void
_cache_sort_vnodes(void ** p1,void ** p2)1008 _cache_sort_vnodes(void **p1, void **p2)
1009 {
1010 	void *tmp;
1011 
1012 	MPASS(*p1 != NULL || *p2 != NULL);
1013 
1014 	if (*p1 > *p2) {
1015 		tmp = *p2;
1016 		*p2 = *p1;
1017 		*p1 = tmp;
1018 	}
1019 }
1020 
1021 static void
cache_lock_all_buckets(void)1022 cache_lock_all_buckets(void)
1023 {
1024 	u_int i;
1025 
1026 	for (i = 0; i < numbucketlocks; i++)
1027 		mtx_lock(&bucketlocks[i]);
1028 }
1029 
1030 static void
cache_unlock_all_buckets(void)1031 cache_unlock_all_buckets(void)
1032 {
1033 	u_int i;
1034 
1035 	for (i = 0; i < numbucketlocks; i++)
1036 		mtx_unlock(&bucketlocks[i]);
1037 }
1038 
1039 static void
cache_lock_all_vnodes(void)1040 cache_lock_all_vnodes(void)
1041 {
1042 	u_int i;
1043 
1044 	for (i = 0; i < numvnodelocks; i++)
1045 		mtx_lock(&vnodelocks[i]);
1046 }
1047 
1048 static void
cache_unlock_all_vnodes(void)1049 cache_unlock_all_vnodes(void)
1050 {
1051 	u_int i;
1052 
1053 	for (i = 0; i < numvnodelocks; i++)
1054 		mtx_unlock(&vnodelocks[i]);
1055 }
1056 
1057 static int
cache_trylock_vnodes(struct mtx * vlp1,struct mtx * vlp2)1058 cache_trylock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
1059 {
1060 
1061 	cache_sort_vnodes(&vlp1, &vlp2);
1062 
1063 	if (vlp1 != NULL) {
1064 		if (!mtx_trylock(vlp1))
1065 			return (EAGAIN);
1066 	}
1067 	if (!mtx_trylock(vlp2)) {
1068 		if (vlp1 != NULL)
1069 			mtx_unlock(vlp1);
1070 		return (EAGAIN);
1071 	}
1072 
1073 	return (0);
1074 }
1075 
1076 static void
cache_lock_vnodes(struct mtx * vlp1,struct mtx * vlp2)1077 cache_lock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
1078 {
1079 
1080 	MPASS(vlp1 != NULL || vlp2 != NULL);
1081 	MPASS(vlp1 <= vlp2);
1082 
1083 	if (vlp1 != NULL)
1084 		mtx_lock(vlp1);
1085 	if (vlp2 != NULL)
1086 		mtx_lock(vlp2);
1087 }
1088 
1089 static void
cache_unlock_vnodes(struct mtx * vlp1,struct mtx * vlp2)1090 cache_unlock_vnodes(struct mtx *vlp1, struct mtx *vlp2)
1091 {
1092 
1093 	MPASS(vlp1 != NULL || vlp2 != NULL);
1094 
1095 	if (vlp1 != NULL)
1096 		mtx_unlock(vlp1);
1097 	if (vlp2 != NULL)
1098 		mtx_unlock(vlp2);
1099 }
1100 
1101 static int
sysctl_nchstats(SYSCTL_HANDLER_ARGS)1102 sysctl_nchstats(SYSCTL_HANDLER_ARGS)
1103 {
1104 	struct nchstats snap;
1105 
1106 	if (req->oldptr == NULL)
1107 		return (SYSCTL_OUT(req, 0, sizeof(snap)));
1108 
1109 	snap = nchstats;
1110 	snap.ncs_goodhits = counter_u64_fetch(numposhits);
1111 	snap.ncs_neghits = counter_u64_fetch(numneghits);
1112 	snap.ncs_badhits = counter_u64_fetch(numposzaps) +
1113 	    counter_u64_fetch(numnegzaps);
1114 	snap.ncs_miss = counter_u64_fetch(nummisszap) +
1115 	    counter_u64_fetch(nummiss);
1116 
1117 	return (SYSCTL_OUT(req, &snap, sizeof(snap)));
1118 }
1119 SYSCTL_PROC(_vfs_cache, OID_AUTO, nchstats, CTLTYPE_OPAQUE | CTLFLAG_RD |
1120     CTLFLAG_MPSAFE, 0, 0, sysctl_nchstats, "LU",
1121     "VFS cache effectiveness statistics");
1122 
1123 static int
sysctl_hitpct(SYSCTL_HANDLER_ARGS)1124 sysctl_hitpct(SYSCTL_HANDLER_ARGS)
1125 {
1126 	long poshits, neghits, miss, total;
1127 	long pct;
1128 
1129 	poshits = counter_u64_fetch(numposhits);
1130 	neghits = counter_u64_fetch(numneghits);
1131 	miss = counter_u64_fetch(nummiss);
1132 	total = poshits + neghits + miss;
1133 
1134 	pct = 0;
1135 	if (total != 0)
1136 		pct = ((poshits + neghits) * 100) / total;
1137 	return (sysctl_handle_int(oidp, 0, pct, req));
1138 }
1139 SYSCTL_PROC(_vfs_cache_stats, OID_AUTO, hitpct,
1140     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_hitpct,
1141     "I", "Percentage of hits");
1142 
1143 static void
cache_recalc_neg_min(void)1144 cache_recalc_neg_min(void)
1145 {
1146 
1147 	neg_min = (ncsize * ncnegminpct) / 100;
1148 }
1149 
1150 static int
sysctl_negminpct(SYSCTL_HANDLER_ARGS)1151 sysctl_negminpct(SYSCTL_HANDLER_ARGS)
1152 {
1153 	u_int val;
1154 	int error;
1155 
1156 	val = ncnegminpct;
1157 	error = sysctl_handle_int(oidp, &val, 0, req);
1158 	if (error != 0 || req->newptr == NULL)
1159 		return (error);
1160 
1161 	if (val == ncnegminpct)
1162 		return (0);
1163 	if (val < 0 || val > 99)
1164 		return (EINVAL);
1165 	ncnegminpct = val;
1166 	cache_recalc_neg_min();
1167 	return (0);
1168 }
1169 
1170 SYSCTL_PROC(_vfs_cache_param, OID_AUTO, negminpct,
1171     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_negminpct,
1172     "I", "Negative entry \% of namecache capacity above which automatic eviction is allowed");
1173 
1174 #ifdef DEBUG_CACHE
1175 /*
1176  * Grab an atomic snapshot of the name cache hash chain lengths
1177  */
1178 static SYSCTL_NODE(_debug, OID_AUTO, hashstat,
1179     CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
1180     "hash table stats");
1181 
1182 static int
sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)1183 sysctl_debug_hashstat_rawnchash(SYSCTL_HANDLER_ARGS)
1184 {
1185 	struct nchashhead *ncpp;
1186 	struct namecache *ncp;
1187 	int i, error, n_nchash, *cntbuf;
1188 
1189 retry:
1190 	n_nchash = nchash + 1;	/* nchash is max index, not count */
1191 	if (req->oldptr == NULL)
1192 		return SYSCTL_OUT(req, 0, n_nchash * sizeof(int));
1193 	cntbuf = malloc(n_nchash * sizeof(int), M_TEMP, M_ZERO | M_WAITOK);
1194 	cache_lock_all_buckets();
1195 	if (n_nchash != nchash + 1) {
1196 		cache_unlock_all_buckets();
1197 		free(cntbuf, M_TEMP);
1198 		goto retry;
1199 	}
1200 	/* Scan hash tables counting entries */
1201 	for (ncpp = nchashtbl, i = 0; i < n_nchash; ncpp++, i++)
1202 		CK_SLIST_FOREACH(ncp, ncpp, nc_hash)
1203 			cntbuf[i]++;
1204 	cache_unlock_all_buckets();
1205 	for (error = 0, i = 0; i < n_nchash; i++)
1206 		if ((error = SYSCTL_OUT(req, &cntbuf[i], sizeof(int))) != 0)
1207 			break;
1208 	free(cntbuf, M_TEMP);
1209 	return (error);
1210 }
1211 SYSCTL_PROC(_debug_hashstat, OID_AUTO, rawnchash, CTLTYPE_INT|CTLFLAG_RD|
1212     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_rawnchash, "S,int",
1213     "nchash chain lengths");
1214 
1215 static int
sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)1216 sysctl_debug_hashstat_nchash(SYSCTL_HANDLER_ARGS)
1217 {
1218 	int error;
1219 	struct nchashhead *ncpp;
1220 	struct namecache *ncp;
1221 	int n_nchash;
1222 	int count, maxlength, used, pct;
1223 
1224 	if (!req->oldptr)
1225 		return SYSCTL_OUT(req, 0, 4 * sizeof(int));
1226 
1227 	cache_lock_all_buckets();
1228 	n_nchash = nchash + 1;	/* nchash is max index, not count */
1229 	used = 0;
1230 	maxlength = 0;
1231 
1232 	/* Scan hash tables for applicable entries */
1233 	for (ncpp = nchashtbl; n_nchash > 0; n_nchash--, ncpp++) {
1234 		count = 0;
1235 		CK_SLIST_FOREACH(ncp, ncpp, nc_hash) {
1236 			count++;
1237 		}
1238 		if (count)
1239 			used++;
1240 		if (maxlength < count)
1241 			maxlength = count;
1242 	}
1243 	n_nchash = nchash + 1;
1244 	cache_unlock_all_buckets();
1245 	pct = (used * 100) / (n_nchash / 100);
1246 	error = SYSCTL_OUT(req, &n_nchash, sizeof(n_nchash));
1247 	if (error)
1248 		return (error);
1249 	error = SYSCTL_OUT(req, &used, sizeof(used));
1250 	if (error)
1251 		return (error);
1252 	error = SYSCTL_OUT(req, &maxlength, sizeof(maxlength));
1253 	if (error)
1254 		return (error);
1255 	error = SYSCTL_OUT(req, &pct, sizeof(pct));
1256 	if (error)
1257 		return (error);
1258 	return (0);
1259 }
1260 SYSCTL_PROC(_debug_hashstat, OID_AUTO, nchash, CTLTYPE_INT|CTLFLAG_RD|
1261     CTLFLAG_MPSAFE, 0, 0, sysctl_debug_hashstat_nchash, "I",
1262     "nchash statistics (number of total/used buckets, maximum chain length, usage percentage)");
1263 #endif
1264 
1265 /*
1266  * Negative entries management
1267  *
1268  * Various workloads create plenty of negative entries and barely use them
1269  * afterwards. Moreover malicious users can keep performing bogus lookups
1270  * adding even more entries. For example "make tinderbox" as of writing this
1271  * comment ends up with 2.6M namecache entries in total, 1.2M of which are
1272  * negative.
1273  *
1274  * As such, a rather aggressive eviction method is needed. The currently
1275  * employed method is a placeholder.
1276  *
1277  * Entries are split over numneglists separate lists, each of which is further
1278  * split into hot and cold entries. Entries get promoted after getting a hit.
1279  * Eviction happens on addition of new entry.
1280  */
1281 static SYSCTL_NODE(_vfs_cache, OID_AUTO, neg, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1282     "Name cache negative entry statistics");
1283 
1284 SYSCTL_ULONG(_vfs_cache_neg, OID_AUTO, count, CTLFLAG_RD, &numneg, 0,
1285     "Number of negative cache entries");
1286 
1287 static COUNTER_U64_DEFINE_EARLY(neg_created);
1288 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, created, CTLFLAG_RD, &neg_created,
1289     "Number of created negative entries");
1290 
1291 static COUNTER_U64_DEFINE_EARLY(neg_evicted);
1292 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evicted, CTLFLAG_RD, &neg_evicted,
1293     "Number of evicted negative entries");
1294 
1295 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_empty);
1296 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_empty, CTLFLAG_RD,
1297     &neg_evict_skipped_empty,
1298     "Number of times evicting failed due to lack of entries");
1299 
1300 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_missed);
1301 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_missed, CTLFLAG_RD,
1302     &neg_evict_skipped_missed,
1303     "Number of times evicting failed due to target entry disappearing");
1304 
1305 static COUNTER_U64_DEFINE_EARLY(neg_evict_skipped_contended);
1306 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, evict_skipped_contended, CTLFLAG_RD,
1307     &neg_evict_skipped_contended,
1308     "Number of times evicting failed due to contention");
1309 
1310 SYSCTL_COUNTER_U64(_vfs_cache_neg, OID_AUTO, hits, CTLFLAG_RD, &numneghits,
1311     "Number of cache hits (negative)");
1312 
1313 static int
sysctl_neg_hot(SYSCTL_HANDLER_ARGS)1314 sysctl_neg_hot(SYSCTL_HANDLER_ARGS)
1315 {
1316 	int i, out;
1317 
1318 	out = 0;
1319 	for (i = 0; i < numneglists; i++)
1320 		out += neglists[i].nl_hotnum;
1321 
1322 	return (SYSCTL_OUT(req, &out, sizeof(out)));
1323 }
1324 SYSCTL_PROC(_vfs_cache_neg, OID_AUTO, hot, CTLTYPE_INT | CTLFLAG_RD |
1325     CTLFLAG_MPSAFE, 0, 0, sysctl_neg_hot, "I",
1326     "Number of hot negative entries");
1327 
1328 static void
cache_neg_init(struct namecache * ncp)1329 cache_neg_init(struct namecache *ncp)
1330 {
1331 	struct negstate *ns;
1332 
1333 	ncp->nc_flag |= NCF_NEGATIVE;
1334 	ns = NCP2NEGSTATE(ncp);
1335 	ns->neg_flag = 0;
1336 	ns->neg_hit = 0;
1337 	counter_u64_add(neg_created, 1);
1338 }
1339 
1340 #define CACHE_NEG_PROMOTION_THRESH 2
1341 
1342 static bool
cache_neg_hit_prep(struct namecache * ncp)1343 cache_neg_hit_prep(struct namecache *ncp)
1344 {
1345 	struct negstate *ns;
1346 	u_char n;
1347 
1348 	ns = NCP2NEGSTATE(ncp);
1349 	n = atomic_load_char(&ns->neg_hit);
1350 	for (;;) {
1351 		if (n >= CACHE_NEG_PROMOTION_THRESH)
1352 			return (false);
1353 		if (atomic_fcmpset_8(&ns->neg_hit, &n, n + 1))
1354 			break;
1355 	}
1356 	return (n + 1 == CACHE_NEG_PROMOTION_THRESH);
1357 }
1358 
1359 /*
1360  * Nothing to do here but it is provided for completeness as some
1361  * cache_neg_hit_prep callers may end up returning without even
1362  * trying to promote.
1363  */
1364 #define cache_neg_hit_abort(ncp)	do { } while (0)
1365 
1366 static void
cache_neg_hit_finish(struct namecache * ncp)1367 cache_neg_hit_finish(struct namecache *ncp)
1368 {
1369 
1370 	SDT_PROBE2(vfs, namecache, lookup, hit__negative, ncp->nc_dvp, ncp->nc_name);
1371 	counter_u64_add(numneghits, 1);
1372 }
1373 
1374 /*
1375  * Move a negative entry to the hot list.
1376  */
1377 static void
cache_neg_promote_locked(struct namecache * ncp)1378 cache_neg_promote_locked(struct namecache *ncp)
1379 {
1380 	struct neglist *nl;
1381 	struct negstate *ns;
1382 
1383 	ns = NCP2NEGSTATE(ncp);
1384 	nl = NCP2NEGLIST(ncp);
1385 	mtx_assert(&nl->nl_lock, MA_OWNED);
1386 	if ((ns->neg_flag & NEG_HOT) == 0) {
1387 		TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1388 		TAILQ_INSERT_TAIL(&nl->nl_hotlist, ncp, nc_dst);
1389 		nl->nl_hotnum++;
1390 		ns->neg_flag |= NEG_HOT;
1391 	}
1392 }
1393 
1394 /*
1395  * Move a hot negative entry to the cold list.
1396  */
1397 static void
cache_neg_demote_locked(struct namecache * ncp)1398 cache_neg_demote_locked(struct namecache *ncp)
1399 {
1400 	struct neglist *nl;
1401 	struct negstate *ns;
1402 
1403 	ns = NCP2NEGSTATE(ncp);
1404 	nl = NCP2NEGLIST(ncp);
1405 	mtx_assert(&nl->nl_lock, MA_OWNED);
1406 	MPASS(ns->neg_flag & NEG_HOT);
1407 	TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1408 	TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1409 	nl->nl_hotnum--;
1410 	ns->neg_flag &= ~NEG_HOT;
1411 	atomic_store_char(&ns->neg_hit, 0);
1412 }
1413 
1414 /*
1415  * Move a negative entry to the hot list if it matches the lookup.
1416  *
1417  * We have to take locks, but they may be contended and in the worst
1418  * case we may need to go off CPU. We don't want to spin within the
1419  * smr section and we can't block with it. Exiting the section means
1420  * the found entry could have been evicted. We are going to look it
1421  * up again.
1422  */
1423 static bool
cache_neg_promote_cond(struct vnode * dvp,struct componentname * cnp,struct namecache * oncp,uint32_t hash)1424 cache_neg_promote_cond(struct vnode *dvp, struct componentname *cnp,
1425     struct namecache *oncp, uint32_t hash)
1426 {
1427 	struct namecache *ncp;
1428 	struct neglist *nl;
1429 	u_char nc_flag;
1430 
1431 	nl = NCP2NEGLIST(oncp);
1432 
1433 	mtx_lock(&nl->nl_lock);
1434 	/*
1435 	 * For hash iteration.
1436 	 */
1437 	vfs_smr_enter();
1438 
1439 	/*
1440 	 * Avoid all surprises by only succeeding if we got the same entry and
1441 	 * bailing completely otherwise.
1442 	 * XXX There are no provisions to keep the vnode around, meaning we may
1443 	 * end up promoting a negative entry for a *new* vnode and returning
1444 	 * ENOENT on its account. This is the error we want to return anyway
1445 	 * and promotion is harmless.
1446 	 *
1447 	 * In particular at this point there can be a new ncp which matches the
1448 	 * search but hashes to a different neglist.
1449 	 */
1450 	CK_SLIST_FOREACH(ncp, (NCHHASH(hash)), nc_hash) {
1451 		if (ncp == oncp)
1452 			break;
1453 	}
1454 
1455 	/*
1456 	 * No match to begin with.
1457 	 */
1458 	if (__predict_false(ncp == NULL)) {
1459 		goto out_abort;
1460 	}
1461 
1462 	/*
1463 	 * The newly found entry may be something different...
1464 	 */
1465 	if (!cache_ncp_match(ncp, dvp, cnp)) {
1466 		goto out_abort;
1467 	}
1468 
1469 	/*
1470 	 * ... and not even negative.
1471 	 */
1472 	nc_flag = atomic_load_char(&ncp->nc_flag);
1473 	if ((nc_flag & NCF_NEGATIVE) == 0) {
1474 		goto out_abort;
1475 	}
1476 
1477 	if (!cache_ncp_canuse(ncp)) {
1478 		goto out_abort;
1479 	}
1480 
1481 	cache_neg_promote_locked(ncp);
1482 	cache_neg_hit_finish(ncp);
1483 	vfs_smr_exit();
1484 	mtx_unlock(&nl->nl_lock);
1485 	return (true);
1486 out_abort:
1487 	vfs_smr_exit();
1488 	mtx_unlock(&nl->nl_lock);
1489 	return (false);
1490 }
1491 
1492 static void
cache_neg_promote(struct namecache * ncp)1493 cache_neg_promote(struct namecache *ncp)
1494 {
1495 	struct neglist *nl;
1496 
1497 	nl = NCP2NEGLIST(ncp);
1498 	mtx_lock(&nl->nl_lock);
1499 	cache_neg_promote_locked(ncp);
1500 	mtx_unlock(&nl->nl_lock);
1501 }
1502 
1503 static void
cache_neg_insert(struct namecache * ncp)1504 cache_neg_insert(struct namecache *ncp)
1505 {
1506 	struct neglist *nl;
1507 
1508 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
1509 	cache_assert_bucket_locked(ncp);
1510 	nl = NCP2NEGLIST(ncp);
1511 	mtx_lock(&nl->nl_lock);
1512 	TAILQ_INSERT_TAIL(&nl->nl_list, ncp, nc_dst);
1513 	mtx_unlock(&nl->nl_lock);
1514 	atomic_add_long(&numneg, 1);
1515 }
1516 
1517 static void
cache_neg_remove(struct namecache * ncp)1518 cache_neg_remove(struct namecache *ncp)
1519 {
1520 	struct neglist *nl;
1521 	struct negstate *ns;
1522 
1523 	cache_assert_bucket_locked(ncp);
1524 	nl = NCP2NEGLIST(ncp);
1525 	ns = NCP2NEGSTATE(ncp);
1526 	mtx_lock(&nl->nl_lock);
1527 	if ((ns->neg_flag & NEG_HOT) != 0) {
1528 		TAILQ_REMOVE(&nl->nl_hotlist, ncp, nc_dst);
1529 		nl->nl_hotnum--;
1530 	} else {
1531 		TAILQ_REMOVE(&nl->nl_list, ncp, nc_dst);
1532 	}
1533 	mtx_unlock(&nl->nl_lock);
1534 	atomic_subtract_long(&numneg, 1);
1535 }
1536 
1537 static struct neglist *
cache_neg_evict_select_list(void)1538 cache_neg_evict_select_list(void)
1539 {
1540 	struct neglist *nl;
1541 	u_int c;
1542 
1543 	c = atomic_fetchadd_int(&neg_cycle, 1) + 1;
1544 	nl = &neglists[c % numneglists];
1545 	if (!mtx_trylock(&nl->nl_evict_lock)) {
1546 		counter_u64_add(neg_evict_skipped_contended, 1);
1547 		return (NULL);
1548 	}
1549 	return (nl);
1550 }
1551 
1552 static struct namecache *
cache_neg_evict_select_entry(struct neglist * nl)1553 cache_neg_evict_select_entry(struct neglist *nl)
1554 {
1555 	struct namecache *ncp, *lncp;
1556 	struct negstate *ns, *lns;
1557 	int i;
1558 
1559 	mtx_assert(&nl->nl_evict_lock, MA_OWNED);
1560 	mtx_assert(&nl->nl_lock, MA_OWNED);
1561 	ncp = TAILQ_FIRST(&nl->nl_list);
1562 	if (ncp == NULL)
1563 		return (NULL);
1564 	lncp = ncp;
1565 	lns = NCP2NEGSTATE(lncp);
1566 	for (i = 1; i < 4; i++) {
1567 		ncp = TAILQ_NEXT(ncp, nc_dst);
1568 		if (ncp == NULL)
1569 			break;
1570 		ns = NCP2NEGSTATE(ncp);
1571 		if (ns->neg_hit < lns->neg_hit) {
1572 			lncp = ncp;
1573 			lns = ns;
1574 		}
1575 	}
1576 	return (lncp);
1577 }
1578 
1579 static bool
cache_neg_evict(void)1580 cache_neg_evict(void)
1581 {
1582 	struct namecache *ncp, *ncp2;
1583 	struct neglist *nl;
1584 	struct vnode *dvp;
1585 	struct mtx *dvlp;
1586 	struct mtx *blp;
1587 	uint32_t hash;
1588 	u_char nlen;
1589 	bool evicted;
1590 
1591 	nl = cache_neg_evict_select_list();
1592 	if (nl == NULL) {
1593 		return (false);
1594 	}
1595 
1596 	mtx_lock(&nl->nl_lock);
1597 	ncp = TAILQ_FIRST(&nl->nl_hotlist);
1598 	if (ncp != NULL) {
1599 		cache_neg_demote_locked(ncp);
1600 	}
1601 	ncp = cache_neg_evict_select_entry(nl);
1602 	if (ncp == NULL) {
1603 		counter_u64_add(neg_evict_skipped_empty, 1);
1604 		mtx_unlock(&nl->nl_lock);
1605 		mtx_unlock(&nl->nl_evict_lock);
1606 		return (false);
1607 	}
1608 	nlen = ncp->nc_nlen;
1609 	dvp = ncp->nc_dvp;
1610 	hash = cache_get_hash(ncp->nc_name, nlen, dvp);
1611 	dvlp = VP2VNODELOCK(dvp);
1612 	blp = HASH2BUCKETLOCK(hash);
1613 	mtx_unlock(&nl->nl_lock);
1614 	mtx_unlock(&nl->nl_evict_lock);
1615 	mtx_lock(dvlp);
1616 	mtx_lock(blp);
1617 	/*
1618 	 * Note that since all locks were dropped above, the entry may be
1619 	 * gone or reallocated to be something else.
1620 	 */
1621 	CK_SLIST_FOREACH(ncp2, (NCHHASH(hash)), nc_hash) {
1622 		if (ncp2 == ncp && ncp2->nc_dvp == dvp &&
1623 		    ncp2->nc_nlen == nlen && (ncp2->nc_flag & NCF_NEGATIVE) != 0)
1624 			break;
1625 	}
1626 	if (ncp2 == NULL) {
1627 		counter_u64_add(neg_evict_skipped_missed, 1);
1628 		ncp = NULL;
1629 		evicted = false;
1630 	} else {
1631 		MPASS(dvlp == VP2VNODELOCK(ncp->nc_dvp));
1632 		MPASS(blp == NCP2BUCKETLOCK(ncp));
1633 		SDT_PROBE2(vfs, namecache, evict_negative, done, ncp->nc_dvp,
1634 		    ncp->nc_name);
1635 		cache_zap_locked(ncp);
1636 		counter_u64_add(neg_evicted, 1);
1637 		evicted = true;
1638 	}
1639 	mtx_unlock(blp);
1640 	mtx_unlock(dvlp);
1641 	if (ncp != NULL)
1642 		cache_free(ncp);
1643 	return (evicted);
1644 }
1645 
1646 /*
1647  * Maybe evict a negative entry to create more room.
1648  *
1649  * The ncnegfactor parameter limits what fraction of the total count
1650  * can comprise of negative entries. However, if the cache is just
1651  * warming up this leads to excessive evictions.  As such, ncnegminpct
1652  * (recomputed to neg_min) dictates whether the above should be
1653  * applied.
1654  *
1655  * Try evicting if the cache is close to full capacity regardless of
1656  * other considerations.
1657  */
1658 static bool
cache_neg_evict_cond(u_long lnumcache)1659 cache_neg_evict_cond(u_long lnumcache)
1660 {
1661 	u_long lnumneg;
1662 
1663 	if (ncsize - 1000 < lnumcache)
1664 		goto out_evict;
1665 	lnumneg = atomic_load_long(&numneg);
1666 	if (lnumneg < neg_min)
1667 		return (false);
1668 	if (lnumneg * ncnegfactor < lnumcache)
1669 		return (false);
1670 out_evict:
1671 	return (cache_neg_evict());
1672 }
1673 
1674 /*
1675  * cache_zap_locked():
1676  *
1677  *   Removes a namecache entry from cache, whether it contains an actual
1678  *   pointer to a vnode or if it is just a negative cache entry.
1679  */
1680 static void
cache_zap_locked(struct namecache * ncp)1681 cache_zap_locked(struct namecache *ncp)
1682 {
1683 	struct nchashhead *ncpp;
1684 	struct vnode *dvp, *vp;
1685 
1686 	dvp = ncp->nc_dvp;
1687 	vp = ncp->nc_vp;
1688 
1689 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1690 		cache_assert_vnode_locked(vp);
1691 	cache_assert_vnode_locked(dvp);
1692 	cache_assert_bucket_locked(ncp);
1693 
1694 	cache_ncp_invalidate(ncp);
1695 
1696 	ncpp = NCP2BUCKET(ncp);
1697 	CK_SLIST_REMOVE(ncpp, ncp, namecache, nc_hash);
1698 	if (!(ncp->nc_flag & NCF_NEGATIVE)) {
1699 		SDT_PROBE3(vfs, namecache, zap, done, dvp, ncp->nc_name, vp);
1700 		TAILQ_REMOVE(&vp->v_cache_dst, ncp, nc_dst);
1701 		if (ncp == vp->v_cache_dd) {
1702 			atomic_store_ptr(&vp->v_cache_dd, NULL);
1703 		}
1704 	} else {
1705 		SDT_PROBE2(vfs, namecache, zap_negative, done, dvp, ncp->nc_name);
1706 		cache_neg_remove(ncp);
1707 	}
1708 	if (ncp->nc_flag & NCF_ISDOTDOT) {
1709 		if (ncp == dvp->v_cache_dd) {
1710 			atomic_store_ptr(&dvp->v_cache_dd, NULL);
1711 		}
1712 	} else {
1713 		LIST_REMOVE(ncp, nc_src);
1714 		if (LIST_EMPTY(&dvp->v_cache_src)) {
1715 			ncp->nc_flag |= NCF_DVDROP;
1716 		}
1717 	}
1718 }
1719 
1720 static void
cache_zap_negative_locked_vnode_kl(struct namecache * ncp,struct vnode * vp)1721 cache_zap_negative_locked_vnode_kl(struct namecache *ncp, struct vnode *vp)
1722 {
1723 	struct mtx *blp;
1724 
1725 	MPASS(ncp->nc_dvp == vp);
1726 	MPASS(ncp->nc_flag & NCF_NEGATIVE);
1727 	cache_assert_vnode_locked(vp);
1728 
1729 	blp = NCP2BUCKETLOCK(ncp);
1730 	mtx_lock(blp);
1731 	cache_zap_locked(ncp);
1732 	mtx_unlock(blp);
1733 }
1734 
1735 static bool
cache_zap_locked_vnode_kl2(struct namecache * ncp,struct vnode * vp,struct mtx ** vlpp)1736 cache_zap_locked_vnode_kl2(struct namecache *ncp, struct vnode *vp,
1737     struct mtx **vlpp)
1738 {
1739 	struct mtx *pvlp, *vlp1, *vlp2, *to_unlock;
1740 	struct mtx *blp;
1741 
1742 	MPASS(vp == ncp->nc_dvp || vp == ncp->nc_vp);
1743 	cache_assert_vnode_locked(vp);
1744 
1745 	if (ncp->nc_flag & NCF_NEGATIVE) {
1746 		if (*vlpp != NULL) {
1747 			mtx_unlock(*vlpp);
1748 			*vlpp = NULL;
1749 		}
1750 		cache_zap_negative_locked_vnode_kl(ncp, vp);
1751 		return (true);
1752 	}
1753 
1754 	pvlp = VP2VNODELOCK(vp);
1755 	blp = NCP2BUCKETLOCK(ncp);
1756 	vlp1 = VP2VNODELOCK(ncp->nc_dvp);
1757 	vlp2 = VP2VNODELOCK(ncp->nc_vp);
1758 
1759 	if (*vlpp == vlp1 || *vlpp == vlp2) {
1760 		to_unlock = *vlpp;
1761 		*vlpp = NULL;
1762 	} else {
1763 		if (*vlpp != NULL) {
1764 			mtx_unlock(*vlpp);
1765 			*vlpp = NULL;
1766 		}
1767 		cache_sort_vnodes(&vlp1, &vlp2);
1768 		if (vlp1 == pvlp) {
1769 			mtx_lock(vlp2);
1770 			to_unlock = vlp2;
1771 		} else {
1772 			if (!mtx_trylock(vlp1))
1773 				goto out_relock;
1774 			to_unlock = vlp1;
1775 		}
1776 	}
1777 	mtx_lock(blp);
1778 	cache_zap_locked(ncp);
1779 	mtx_unlock(blp);
1780 	if (to_unlock != NULL)
1781 		mtx_unlock(to_unlock);
1782 	return (true);
1783 
1784 out_relock:
1785 	mtx_unlock(vlp2);
1786 	mtx_lock(vlp1);
1787 	mtx_lock(vlp2);
1788 	MPASS(*vlpp == NULL);
1789 	*vlpp = vlp1;
1790 	return (false);
1791 }
1792 
1793 /*
1794  * If trylocking failed we can get here. We know enough to take all needed locks
1795  * in the right order and re-lookup the entry.
1796  */
1797 static int
cache_zap_unlocked_bucket(struct namecache * ncp,struct componentname * cnp,struct vnode * dvp,struct mtx * dvlp,struct mtx * vlp,uint32_t hash,struct mtx * blp)1798 cache_zap_unlocked_bucket(struct namecache *ncp, struct componentname *cnp,
1799     struct vnode *dvp, struct mtx *dvlp, struct mtx *vlp, uint32_t hash,
1800     struct mtx *blp)
1801 {
1802 	struct namecache *rncp;
1803 	struct mtx *rvlp;
1804 
1805 	cache_assert_bucket_unlocked(ncp);
1806 
1807 	cache_sort_vnodes(&dvlp, &vlp);
1808 	cache_lock_vnodes(dvlp, vlp);
1809 	mtx_lock(blp);
1810 	CK_SLIST_FOREACH(rncp, (NCHHASH(hash)), nc_hash) {
1811 		if (rncp == ncp && cache_ncp_match(rncp, dvp, cnp))
1812 			break;
1813 	}
1814 	if (rncp == NULL)
1815 		goto out_mismatch;
1816 
1817 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1818 		rvlp = VP2VNODELOCK(rncp->nc_vp);
1819 	else
1820 		rvlp = NULL;
1821 	if (rvlp != vlp)
1822 		goto out_mismatch;
1823 
1824 	cache_zap_locked(rncp);
1825 	mtx_unlock(blp);
1826 	cache_unlock_vnodes(dvlp, vlp);
1827 	atomic_add_long(&zap_bucket_relock_success, 1);
1828 	return (0);
1829 
1830 out_mismatch:
1831 	mtx_unlock(blp);
1832 	cache_unlock_vnodes(dvlp, vlp);
1833 	return (EAGAIN);
1834 }
1835 
1836 static int __noinline
cache_zap_locked_bucket(struct namecache * ncp,struct componentname * cnp,uint32_t hash,struct mtx * blp)1837 cache_zap_locked_bucket(struct namecache *ncp, struct componentname *cnp,
1838     uint32_t hash, struct mtx *blp)
1839 {
1840 	struct mtx *dvlp, *vlp;
1841 	struct vnode *dvp;
1842 
1843 	cache_assert_bucket_locked(ncp);
1844 
1845 	dvlp = VP2VNODELOCK(ncp->nc_dvp);
1846 	vlp = NULL;
1847 	if (!(ncp->nc_flag & NCF_NEGATIVE))
1848 		vlp = VP2VNODELOCK(ncp->nc_vp);
1849 	if (cache_trylock_vnodes(dvlp, vlp) == 0) {
1850 		cache_zap_locked(ncp);
1851 		mtx_unlock(blp);
1852 		cache_unlock_vnodes(dvlp, vlp);
1853 		return (0);
1854 	}
1855 
1856 	dvp = ncp->nc_dvp;
1857 	mtx_unlock(blp);
1858 	return (cache_zap_unlocked_bucket(ncp, cnp, dvp, dvlp, vlp, hash, blp));
1859 }
1860 
1861 static __noinline int
cache_remove_cnp(struct vnode * dvp,struct componentname * cnp)1862 cache_remove_cnp(struct vnode *dvp, struct componentname *cnp)
1863 {
1864 	struct namecache *ncp;
1865 	struct mtx *blp;
1866 	struct mtx *dvlp, *dvlp2;
1867 	uint32_t hash;
1868 	int error;
1869 
1870 	if (cnp->cn_namelen == 2 &&
1871 	    cnp->cn_nameptr[0] == '.' && cnp->cn_nameptr[1] == '.') {
1872 		dvlp = VP2VNODELOCK(dvp);
1873 		dvlp2 = NULL;
1874 		mtx_lock(dvlp);
1875 retry_dotdot:
1876 		ncp = dvp->v_cache_dd;
1877 		if (ncp == NULL) {
1878 			mtx_unlock(dvlp);
1879 			if (dvlp2 != NULL)
1880 				mtx_unlock(dvlp2);
1881 			SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1882 			return (0);
1883 		}
1884 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
1885 			if (!cache_zap_locked_vnode_kl2(ncp, dvp, &dvlp2))
1886 				goto retry_dotdot;
1887 			MPASS(dvp->v_cache_dd == NULL);
1888 			mtx_unlock(dvlp);
1889 			if (dvlp2 != NULL)
1890 				mtx_unlock(dvlp2);
1891 			cache_free(ncp);
1892 		} else {
1893 			atomic_store_ptr(&dvp->v_cache_dd, NULL);
1894 			mtx_unlock(dvlp);
1895 			if (dvlp2 != NULL)
1896 				mtx_unlock(dvlp2);
1897 		}
1898 		SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1899 		return (1);
1900 	}
1901 
1902 	/*
1903 	 * XXX note that access here is completely unlocked with no provisions
1904 	 * to keep the hash allocated. If one is sufficiently unlucky a
1905 	 * parallel cache resize can reallocate the hash, unmap backing pages
1906 	 * and cause the empty check below to fault.
1907 	 *
1908 	 * Fixing this has epsilon priority, but can be done with no overhead
1909 	 * for this codepath with sufficient effort.
1910 	 */
1911 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
1912 	blp = HASH2BUCKETLOCK(hash);
1913 retry:
1914 	if (CK_SLIST_EMPTY(NCHHASH(hash)))
1915 		goto out_no_entry;
1916 
1917 	mtx_lock(blp);
1918 	ncp = cache_ncp_find(dvp, cnp, hash);
1919 	if (ncp == NULL) {
1920 		mtx_unlock(blp);
1921 		goto out_no_entry;
1922 	}
1923 
1924 	error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
1925 	if (__predict_false(error != 0)) {
1926 		atomic_add_long(&zap_bucket_fail, 1);
1927 		goto retry;
1928 	}
1929 	counter_u64_add(numposzaps, 1);
1930 	SDT_PROBE2(vfs, namecache, removecnp, hit, dvp, cnp);
1931 	cache_free(ncp);
1932 	return (1);
1933 out_no_entry:
1934 	counter_u64_add(nummisszap, 1);
1935 	SDT_PROBE2(vfs, namecache, removecnp, miss, dvp, cnp);
1936 	return (0);
1937 }
1938 
1939 static int __noinline
cache_lookup_dot(struct vnode * dvp,struct vnode ** vpp,struct componentname * cnp,struct timespec * tsp,int * ticksp)1940 cache_lookup_dot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1941     struct timespec *tsp, int *ticksp)
1942 {
1943 	int ltype;
1944 
1945 	*vpp = dvp;
1946 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ".", *vpp);
1947 	if (tsp != NULL)
1948 		timespecclear(tsp);
1949 	if (ticksp != NULL)
1950 		*ticksp = ticks;
1951 	vrefact(*vpp);
1952 	/*
1953 	 * When we lookup "." we still can be asked to lock it
1954 	 * differently...
1955 	 */
1956 	ltype = cnp->cn_lkflags & LK_TYPE_MASK;
1957 	if (ltype != VOP_ISLOCKED(*vpp)) {
1958 		if (ltype == LK_EXCLUSIVE) {
1959 			vn_lock(*vpp, LK_UPGRADE | LK_RETRY);
1960 			if (VN_IS_DOOMED((*vpp))) {
1961 				/* forced unmount */
1962 				vrele(*vpp);
1963 				*vpp = NULL;
1964 				return (ENOENT);
1965 			}
1966 		} else
1967 			vn_lock(*vpp, LK_DOWNGRADE | LK_RETRY);
1968 	}
1969 	return (-1);
1970 }
1971 
1972 static int __noinline
cache_lookup_dotdot(struct vnode * dvp,struct vnode ** vpp,struct componentname * cnp,struct timespec * tsp,int * ticksp)1973 cache_lookup_dotdot(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
1974     struct timespec *tsp, int *ticksp)
1975 {
1976 	struct namecache_ts *ncp_ts;
1977 	struct namecache *ncp;
1978 	struct mtx *dvlp;
1979 	enum vgetstate vs;
1980 	int error, ltype;
1981 	bool whiteout;
1982 
1983 	MPASS((cnp->cn_flags & ISDOTDOT) != 0);
1984 
1985 	if ((cnp->cn_flags & MAKEENTRY) == 0) {
1986 		cache_remove_cnp(dvp, cnp);
1987 		return (0);
1988 	}
1989 
1990 retry:
1991 	dvlp = VP2VNODELOCK(dvp);
1992 	mtx_lock(dvlp);
1993 	ncp = dvp->v_cache_dd;
1994 	if (ncp == NULL) {
1995 		SDT_PROBE2(vfs, namecache, lookup, miss, dvp, "..");
1996 		mtx_unlock(dvlp);
1997 		return (0);
1998 	}
1999 	if ((ncp->nc_flag & NCF_ISDOTDOT) != 0) {
2000 		if (ncp->nc_flag & NCF_NEGATIVE)
2001 			*vpp = NULL;
2002 		else
2003 			*vpp = ncp->nc_vp;
2004 	} else
2005 		*vpp = ncp->nc_dvp;
2006 	if (*vpp == NULL)
2007 		goto negative_success;
2008 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, "..", *vpp);
2009 	cache_out_ts(ncp, tsp, ticksp);
2010 	if ((ncp->nc_flag & (NCF_ISDOTDOT | NCF_DTS)) ==
2011 	    NCF_DTS && tsp != NULL) {
2012 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
2013 		*tsp = ncp_ts->nc_dotdottime;
2014 	}
2015 
2016 	MPASS(dvp != *vpp);
2017 	ltype = VOP_ISLOCKED(dvp);
2018 	VOP_UNLOCK(dvp);
2019 	vs = vget_prep(*vpp);
2020 	mtx_unlock(dvlp);
2021 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
2022 	vn_lock(dvp, ltype | LK_RETRY);
2023 	if (VN_IS_DOOMED(dvp)) {
2024 		if (error == 0)
2025 			vput(*vpp);
2026 		*vpp = NULL;
2027 		return (ENOENT);
2028 	}
2029 	if (error) {
2030 		*vpp = NULL;
2031 		goto retry;
2032 	}
2033 	return (-1);
2034 negative_success:
2035 	if (__predict_false(cnp->cn_nameiop == CREATE)) {
2036 		if (cnp->cn_flags & ISLASTCN) {
2037 			counter_u64_add(numnegzaps, 1);
2038 			cache_zap_negative_locked_vnode_kl(ncp, dvp);
2039 			mtx_unlock(dvlp);
2040 			cache_free(ncp);
2041 			return (0);
2042 		}
2043 	}
2044 
2045 	whiteout = (ncp->nc_flag & NCF_WHITE);
2046 	cache_out_ts(ncp, tsp, ticksp);
2047 	if (cache_neg_hit_prep(ncp))
2048 		cache_neg_promote(ncp);
2049 	else
2050 		cache_neg_hit_finish(ncp);
2051 	mtx_unlock(dvlp);
2052 	if (whiteout)
2053 		cnp->cn_flags |= ISWHITEOUT;
2054 	return (ENOENT);
2055 }
2056 
2057 /**
2058  * Lookup a name in the name cache
2059  *
2060  * # Arguments
2061  *
2062  * - dvp:	Parent directory in which to search.
2063  * - vpp:	Return argument.  Will contain desired vnode on cache hit.
2064  * - cnp:	Parameters of the name search.  The most interesting bits of
2065  *   		the cn_flags field have the following meanings:
2066  *   	- MAKEENTRY:	If clear, free an entry from the cache rather than look
2067  *   			it up.
2068  *   	- ISDOTDOT:	Must be set if and only if cn_nameptr == ".."
2069  * - tsp:	Return storage for cache timestamp.  On a successful (positive
2070  *   		or negative) lookup, tsp will be filled with any timespec that
2071  *   		was stored when this cache entry was created.  However, it will
2072  *   		be clear for "." entries.
2073  * - ticks:	Return storage for alternate cache timestamp.  On a successful
2074  *   		(positive or negative) lookup, it will contain the ticks value
2075  *   		that was current when the cache entry was created, unless cnp
2076  *   		was ".".
2077  *
2078  * Either both tsp and ticks have to be provided or neither of them.
2079  *
2080  * # Returns
2081  *
2082  * - -1:	A positive cache hit.  vpp will contain the desired vnode.
2083  * - ENOENT:	A negative cache hit, or dvp was recycled out from under us due
2084  *		to a forced unmount.  vpp will not be modified.  If the entry
2085  *		is a whiteout, then the ISWHITEOUT flag will be set in
2086  *		cnp->cn_flags.
2087  * - 0:		A cache miss.  vpp will not be modified.
2088  *
2089  * # Locking
2090  *
2091  * On a cache hit, vpp will be returned locked and ref'd.  If we're looking up
2092  * .., dvp is unlocked.  If we're looking up . an extra ref is taken, but the
2093  * lock is not recursively acquired.
2094  */
2095 static int __noinline
cache_lookup_fallback(struct vnode * dvp,struct vnode ** vpp,struct componentname * cnp,struct timespec * tsp,int * ticksp)2096 cache_lookup_fallback(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
2097     struct timespec *tsp, int *ticksp)
2098 {
2099 	struct namecache *ncp;
2100 	struct mtx *blp;
2101 	uint32_t hash;
2102 	enum vgetstate vs;
2103 	int error;
2104 	bool whiteout;
2105 
2106 	MPASS((cnp->cn_flags & ISDOTDOT) == 0);
2107 	MPASS((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) != 0);
2108 
2109 retry:
2110 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
2111 	blp = HASH2BUCKETLOCK(hash);
2112 	mtx_lock(blp);
2113 
2114 	ncp = cache_ncp_find(dvp, cnp, hash);
2115 	if (__predict_false(ncp == NULL)) {
2116 		mtx_unlock(blp);
2117 		SDT_PROBE2(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr);
2118 		counter_u64_add(nummiss, 1);
2119 		return (0);
2120 	}
2121 
2122 	if (ncp->nc_flag & NCF_NEGATIVE)
2123 		goto negative_success;
2124 
2125 	counter_u64_add(numposhits, 1);
2126 	*vpp = ncp->nc_vp;
2127 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
2128 	cache_out_ts(ncp, tsp, ticksp);
2129 	MPASS(dvp != *vpp);
2130 	vs = vget_prep(*vpp);
2131 	mtx_unlock(blp);
2132 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
2133 	if (error) {
2134 		*vpp = NULL;
2135 		goto retry;
2136 	}
2137 	return (-1);
2138 negative_success:
2139 	/*
2140 	 * We don't get here with regular lookup apart from corner cases.
2141 	 */
2142 	if (__predict_true(cnp->cn_nameiop == CREATE)) {
2143 		if (cnp->cn_flags & ISLASTCN) {
2144 			counter_u64_add(numnegzaps, 1);
2145 			error = cache_zap_locked_bucket(ncp, cnp, hash, blp);
2146 			if (__predict_false(error != 0)) {
2147 				atomic_add_long(&zap_bucket_fail2, 1);
2148 				goto retry;
2149 			}
2150 			cache_free(ncp);
2151 			return (0);
2152 		}
2153 	}
2154 
2155 	whiteout = (ncp->nc_flag & NCF_WHITE);
2156 	cache_out_ts(ncp, tsp, ticksp);
2157 	if (cache_neg_hit_prep(ncp))
2158 		cache_neg_promote(ncp);
2159 	else
2160 		cache_neg_hit_finish(ncp);
2161 	mtx_unlock(blp);
2162 	if (whiteout)
2163 		cnp->cn_flags |= ISWHITEOUT;
2164 	return (ENOENT);
2165 }
2166 
2167 int
cache_lookup(struct vnode * dvp,struct vnode ** vpp,struct componentname * cnp,struct timespec * tsp,int * ticksp)2168 cache_lookup(struct vnode *dvp, struct vnode **vpp, struct componentname *cnp,
2169     struct timespec *tsp, int *ticksp)
2170 {
2171 	struct namecache *ncp;
2172 	uint32_t hash;
2173 	enum vgetstate vs;
2174 	int error;
2175 	bool whiteout, neg_promote;
2176 	u_short nc_flag;
2177 
2178 	MPASS((tsp == NULL && ticksp == NULL) || (tsp != NULL && ticksp != NULL));
2179 
2180 #ifdef DEBUG_CACHE
2181 	if (__predict_false(!doingcache)) {
2182 		cnp->cn_flags &= ~MAKEENTRY;
2183 		return (0);
2184 	}
2185 #endif
2186 
2187 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
2188 		if (cnp->cn_namelen == 1)
2189 			return (cache_lookup_dot(dvp, vpp, cnp, tsp, ticksp));
2190 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.')
2191 			return (cache_lookup_dotdot(dvp, vpp, cnp, tsp, ticksp));
2192 	}
2193 
2194 	MPASS((cnp->cn_flags & ISDOTDOT) == 0);
2195 
2196 	if ((cnp->cn_flags & (MAKEENTRY | NC_KEEPPOSENTRY)) == 0) {
2197 		cache_remove_cnp(dvp, cnp);
2198 		return (0);
2199 	}
2200 
2201 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
2202 	vfs_smr_enter();
2203 
2204 	ncp = cache_ncp_find(dvp, cnp, hash);
2205 	if (__predict_false(ncp == NULL)) {
2206 		vfs_smr_exit();
2207 		SDT_PROBE2(vfs, namecache, lookup, miss, dvp, cnp->cn_nameptr);
2208 		counter_u64_add(nummiss, 1);
2209 		return (0);
2210 	}
2211 
2212 	nc_flag = atomic_load_char(&ncp->nc_flag);
2213 	if (nc_flag & NCF_NEGATIVE)
2214 		goto negative_success;
2215 
2216 	counter_u64_add(numposhits, 1);
2217 	*vpp = ncp->nc_vp;
2218 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, *vpp);
2219 	cache_out_ts(ncp, tsp, ticksp);
2220 	MPASS(dvp != *vpp);
2221 	if (!cache_ncp_canuse(ncp)) {
2222 		vfs_smr_exit();
2223 		*vpp = NULL;
2224 		goto out_fallback;
2225 	}
2226 	vs = vget_prep_smr(*vpp);
2227 	vfs_smr_exit();
2228 	if (__predict_false(vs == VGET_NONE)) {
2229 		*vpp = NULL;
2230 		goto out_fallback;
2231 	}
2232 	error = vget_finish(*vpp, cnp->cn_lkflags, vs);
2233 	if (error) {
2234 		*vpp = NULL;
2235 		goto out_fallback;
2236 	}
2237 	return (-1);
2238 negative_success:
2239 	if (cnp->cn_nameiop == CREATE) {
2240 		if (cnp->cn_flags & ISLASTCN) {
2241 			vfs_smr_exit();
2242 			goto out_fallback;
2243 		}
2244 	}
2245 
2246 	cache_out_ts(ncp, tsp, ticksp);
2247 	whiteout = (atomic_load_char(&ncp->nc_flag) & NCF_WHITE);
2248 	neg_promote = cache_neg_hit_prep(ncp);
2249 	if (!cache_ncp_canuse(ncp)) {
2250 		cache_neg_hit_abort(ncp);
2251 		vfs_smr_exit();
2252 		goto out_fallback;
2253 	}
2254 	if (neg_promote) {
2255 		vfs_smr_exit();
2256 		if (!cache_neg_promote_cond(dvp, cnp, ncp, hash))
2257 			goto out_fallback;
2258 	} else {
2259 		cache_neg_hit_finish(ncp);
2260 		vfs_smr_exit();
2261 	}
2262 	if (whiteout)
2263 		cnp->cn_flags |= ISWHITEOUT;
2264 	return (ENOENT);
2265 out_fallback:
2266 	return (cache_lookup_fallback(dvp, vpp, cnp, tsp, ticksp));
2267 }
2268 
2269 struct celockstate {
2270 	struct mtx *vlp[3];
2271 	struct mtx *blp[2];
2272 };
2273 CTASSERT((nitems(((struct celockstate *)0)->vlp) == 3));
2274 CTASSERT((nitems(((struct celockstate *)0)->blp) == 2));
2275 
2276 static inline void
cache_celockstate_init(struct celockstate * cel)2277 cache_celockstate_init(struct celockstate *cel)
2278 {
2279 
2280 	bzero(cel, sizeof(*cel));
2281 }
2282 
2283 static void
cache_lock_vnodes_cel(struct celockstate * cel,struct vnode * vp,struct vnode * dvp)2284 cache_lock_vnodes_cel(struct celockstate *cel, struct vnode *vp,
2285     struct vnode *dvp)
2286 {
2287 	struct mtx *vlp1, *vlp2;
2288 
2289 	MPASS(cel->vlp[0] == NULL);
2290 	MPASS(cel->vlp[1] == NULL);
2291 	MPASS(cel->vlp[2] == NULL);
2292 
2293 	MPASS(vp != NULL || dvp != NULL);
2294 
2295 	vlp1 = VP2VNODELOCK(vp);
2296 	vlp2 = VP2VNODELOCK(dvp);
2297 	cache_sort_vnodes(&vlp1, &vlp2);
2298 
2299 	if (vlp1 != NULL) {
2300 		mtx_lock(vlp1);
2301 		cel->vlp[0] = vlp1;
2302 	}
2303 	mtx_lock(vlp2);
2304 	cel->vlp[1] = vlp2;
2305 }
2306 
2307 static void
cache_unlock_vnodes_cel(struct celockstate * cel)2308 cache_unlock_vnodes_cel(struct celockstate *cel)
2309 {
2310 
2311 	MPASS(cel->vlp[0] != NULL || cel->vlp[1] != NULL);
2312 
2313 	if (cel->vlp[0] != NULL)
2314 		mtx_unlock(cel->vlp[0]);
2315 	if (cel->vlp[1] != NULL)
2316 		mtx_unlock(cel->vlp[1]);
2317 	if (cel->vlp[2] != NULL)
2318 		mtx_unlock(cel->vlp[2]);
2319 }
2320 
2321 static bool
cache_lock_vnodes_cel_3(struct celockstate * cel,struct vnode * vp)2322 cache_lock_vnodes_cel_3(struct celockstate *cel, struct vnode *vp)
2323 {
2324 	struct mtx *vlp;
2325 	bool ret;
2326 
2327 	cache_assert_vlp_locked(cel->vlp[0]);
2328 	cache_assert_vlp_locked(cel->vlp[1]);
2329 	MPASS(cel->vlp[2] == NULL);
2330 
2331 	MPASS(vp != NULL);
2332 	vlp = VP2VNODELOCK(vp);
2333 
2334 	ret = true;
2335 	if (vlp >= cel->vlp[1]) {
2336 		mtx_lock(vlp);
2337 	} else {
2338 		if (mtx_trylock(vlp))
2339 			goto out;
2340 		cache_unlock_vnodes_cel(cel);
2341 		atomic_add_long(&cache_lock_vnodes_cel_3_failures, 1);
2342 		if (vlp < cel->vlp[0]) {
2343 			mtx_lock(vlp);
2344 			mtx_lock(cel->vlp[0]);
2345 			mtx_lock(cel->vlp[1]);
2346 		} else {
2347 			if (cel->vlp[0] != NULL)
2348 				mtx_lock(cel->vlp[0]);
2349 			mtx_lock(vlp);
2350 			mtx_lock(cel->vlp[1]);
2351 		}
2352 		ret = false;
2353 	}
2354 out:
2355 	cel->vlp[2] = vlp;
2356 	return (ret);
2357 }
2358 
2359 static void
cache_lock_buckets_cel(struct celockstate * cel,struct mtx * blp1,struct mtx * blp2)2360 cache_lock_buckets_cel(struct celockstate *cel, struct mtx *blp1,
2361     struct mtx *blp2)
2362 {
2363 
2364 	MPASS(cel->blp[0] == NULL);
2365 	MPASS(cel->blp[1] == NULL);
2366 
2367 	cache_sort_vnodes(&blp1, &blp2);
2368 
2369 	if (blp1 != NULL) {
2370 		mtx_lock(blp1);
2371 		cel->blp[0] = blp1;
2372 	}
2373 	mtx_lock(blp2);
2374 	cel->blp[1] = blp2;
2375 }
2376 
2377 static void
cache_unlock_buckets_cel(struct celockstate * cel)2378 cache_unlock_buckets_cel(struct celockstate *cel)
2379 {
2380 
2381 	if (cel->blp[0] != NULL)
2382 		mtx_unlock(cel->blp[0]);
2383 	mtx_unlock(cel->blp[1]);
2384 }
2385 
2386 /*
2387  * Lock part of the cache affected by the insertion.
2388  *
2389  * This means vnodelocks for dvp, vp and the relevant bucketlock.
2390  * However, insertion can result in removal of an old entry. In this
2391  * case we have an additional vnode and bucketlock pair to lock.
2392  *
2393  * That is, in the worst case we have to lock 3 vnodes and 2 bucketlocks, while
2394  * preserving the locking order (smaller address first).
2395  */
2396 static void
cache_enter_lock(struct celockstate * cel,struct vnode * dvp,struct vnode * vp,uint32_t hash)2397 cache_enter_lock(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2398     uint32_t hash)
2399 {
2400 	struct namecache *ncp;
2401 	struct mtx *blps[2];
2402 	u_char nc_flag;
2403 
2404 	blps[0] = HASH2BUCKETLOCK(hash);
2405 	for (;;) {
2406 		blps[1] = NULL;
2407 		cache_lock_vnodes_cel(cel, dvp, vp);
2408 		if (vp == NULL || vp->v_type != VDIR)
2409 			break;
2410 		ncp = atomic_load_consume_ptr(&vp->v_cache_dd);
2411 		if (ncp == NULL)
2412 			break;
2413 		nc_flag = atomic_load_char(&ncp->nc_flag);
2414 		if ((nc_flag & NCF_ISDOTDOT) == 0)
2415 			break;
2416 		MPASS(ncp->nc_dvp == vp);
2417 		blps[1] = NCP2BUCKETLOCK(ncp);
2418 		if ((nc_flag & NCF_NEGATIVE) != 0)
2419 			break;
2420 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2421 			break;
2422 		/*
2423 		 * All vnodes got re-locked. Re-validate the state and if
2424 		 * nothing changed we are done. Otherwise restart.
2425 		 */
2426 		if (ncp == vp->v_cache_dd &&
2427 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2428 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
2429 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2430 			break;
2431 		cache_unlock_vnodes_cel(cel);
2432 		cel->vlp[0] = NULL;
2433 		cel->vlp[1] = NULL;
2434 		cel->vlp[2] = NULL;
2435 	}
2436 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
2437 }
2438 
2439 static void
cache_enter_lock_dd(struct celockstate * cel,struct vnode * dvp,struct vnode * vp,uint32_t hash)2440 cache_enter_lock_dd(struct celockstate *cel, struct vnode *dvp, struct vnode *vp,
2441     uint32_t hash)
2442 {
2443 	struct namecache *ncp;
2444 	struct mtx *blps[2];
2445 	u_char nc_flag;
2446 
2447 	blps[0] = HASH2BUCKETLOCK(hash);
2448 	for (;;) {
2449 		blps[1] = NULL;
2450 		cache_lock_vnodes_cel(cel, dvp, vp);
2451 		ncp = atomic_load_consume_ptr(&dvp->v_cache_dd);
2452 		if (ncp == NULL)
2453 			break;
2454 		nc_flag = atomic_load_char(&ncp->nc_flag);
2455 		if ((nc_flag & NCF_ISDOTDOT) == 0)
2456 			break;
2457 		MPASS(ncp->nc_dvp == dvp);
2458 		blps[1] = NCP2BUCKETLOCK(ncp);
2459 		if ((nc_flag & NCF_NEGATIVE) != 0)
2460 			break;
2461 		if (cache_lock_vnodes_cel_3(cel, ncp->nc_vp))
2462 			break;
2463 		if (ncp == dvp->v_cache_dd &&
2464 		    (ncp->nc_flag & NCF_ISDOTDOT) != 0 &&
2465 		    blps[1] == NCP2BUCKETLOCK(ncp) &&
2466 		    VP2VNODELOCK(ncp->nc_vp) == cel->vlp[2])
2467 			break;
2468 		cache_unlock_vnodes_cel(cel);
2469 		cel->vlp[0] = NULL;
2470 		cel->vlp[1] = NULL;
2471 		cel->vlp[2] = NULL;
2472 	}
2473 	cache_lock_buckets_cel(cel, blps[0], blps[1]);
2474 }
2475 
2476 static void
cache_enter_unlock(struct celockstate * cel)2477 cache_enter_unlock(struct celockstate *cel)
2478 {
2479 
2480 	cache_unlock_buckets_cel(cel);
2481 	cache_unlock_vnodes_cel(cel);
2482 }
2483 
2484 static void __noinline
cache_enter_dotdot_prep(struct vnode * dvp,struct vnode * vp,struct componentname * cnp)2485 cache_enter_dotdot_prep(struct vnode *dvp, struct vnode *vp,
2486     struct componentname *cnp)
2487 {
2488 	struct celockstate cel;
2489 	struct namecache *ncp;
2490 	uint32_t hash;
2491 	int len;
2492 
2493 	if (atomic_load_ptr(&dvp->v_cache_dd) == NULL)
2494 		return;
2495 	len = cnp->cn_namelen;
2496 	cache_celockstate_init(&cel);
2497 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2498 	cache_enter_lock_dd(&cel, dvp, vp, hash);
2499 	ncp = dvp->v_cache_dd;
2500 	if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT)) {
2501 		KASSERT(ncp->nc_dvp == dvp, ("wrong isdotdot parent"));
2502 		cache_zap_locked(ncp);
2503 	} else {
2504 		ncp = NULL;
2505 	}
2506 	atomic_store_ptr(&dvp->v_cache_dd, NULL);
2507 	cache_enter_unlock(&cel);
2508 	if (ncp != NULL)
2509 		cache_free(ncp);
2510 }
2511 
2512 /*
2513  * Add an entry to the cache.
2514  */
2515 void
cache_enter_time(struct vnode * dvp,struct vnode * vp,struct componentname * cnp,struct timespec * tsp,struct timespec * dtsp)2516 cache_enter_time(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
2517     struct timespec *tsp, struct timespec *dtsp)
2518 {
2519 	struct celockstate cel;
2520 	struct namecache *ncp, *n2, *ndd;
2521 	struct namecache_ts *ncp_ts;
2522 	uint32_t hash;
2523 	int flag;
2524 	int len;
2525 
2526 	KASSERT(cnp->cn_namelen <= NAME_MAX,
2527 	    ("%s: passed len %ld exceeds NAME_MAX (%d)", __func__, cnp->cn_namelen,
2528 	    NAME_MAX));
2529 	VNPASS(!VN_IS_DOOMED(dvp), dvp);
2530 	VNPASS(dvp->v_type != VNON, dvp);
2531 	if (vp != NULL) {
2532 		VNPASS(!VN_IS_DOOMED(vp), vp);
2533 		VNPASS(vp->v_type != VNON, vp);
2534 	}
2535 	if (cnp->cn_namelen == 1 && cnp->cn_nameptr[0] == '.') {
2536 		KASSERT(dvp == vp,
2537 		    ("%s: different vnodes for dot entry (%p; %p)\n", __func__,
2538 		    dvp, vp));
2539 	} else {
2540 		KASSERT(dvp != vp,
2541 		    ("%s: same vnode for non-dot entry [%s] (%p)\n", __func__,
2542 		    cnp->cn_nameptr, dvp));
2543 	}
2544 
2545 #ifdef DEBUG_CACHE
2546 	if (__predict_false(!doingcache))
2547 		return;
2548 #endif
2549 
2550 	flag = 0;
2551 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
2552 		if (cnp->cn_namelen == 1)
2553 			return;
2554 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
2555 			cache_enter_dotdot_prep(dvp, vp, cnp);
2556 			flag = NCF_ISDOTDOT;
2557 		}
2558 	}
2559 
2560 	ncp = cache_alloc(cnp->cn_namelen, tsp != NULL);
2561 	if (ncp == NULL)
2562 		return;
2563 
2564 	cache_celockstate_init(&cel);
2565 	ndd = NULL;
2566 	ncp_ts = NULL;
2567 
2568 	/*
2569 	 * Calculate the hash key and setup as much of the new
2570 	 * namecache entry as possible before acquiring the lock.
2571 	 */
2572 	ncp->nc_flag = flag | NCF_WIP;
2573 	ncp->nc_vp = vp;
2574 	if (vp == NULL)
2575 		cache_neg_init(ncp);
2576 	ncp->nc_dvp = dvp;
2577 	if (tsp != NULL) {
2578 		ncp_ts = __containerof(ncp, struct namecache_ts, nc_nc);
2579 		ncp_ts->nc_time = *tsp;
2580 		ncp_ts->nc_ticks = ticks;
2581 		ncp_ts->nc_nc.nc_flag |= NCF_TS;
2582 		if (dtsp != NULL) {
2583 			ncp_ts->nc_dotdottime = *dtsp;
2584 			ncp_ts->nc_nc.nc_flag |= NCF_DTS;
2585 		}
2586 	}
2587 	len = ncp->nc_nlen = cnp->cn_namelen;
2588 	hash = cache_get_hash(cnp->cn_nameptr, len, dvp);
2589 	memcpy(ncp->nc_name, cnp->cn_nameptr, len);
2590 	ncp->nc_name[len] = '\0';
2591 	cache_enter_lock(&cel, dvp, vp, hash);
2592 
2593 	/*
2594 	 * See if this vnode or negative entry is already in the cache
2595 	 * with this name.  This can happen with concurrent lookups of
2596 	 * the same path name.
2597 	 */
2598 	n2 = cache_ncp_find(dvp, cnp, hash);
2599 	if (n2 != NULL) {
2600 		MPASS(cache_ncp_canuse(n2));
2601 		if ((n2->nc_flag & NCF_NEGATIVE) != 0)
2602 			KASSERT(vp == NULL,
2603 			    ("%s: found entry pointing to a different vnode "
2604 			    "(%p != %p); name [%s]",
2605 			    __func__, NULL, vp, cnp->cn_nameptr));
2606 		else
2607 			KASSERT(n2->nc_vp == vp,
2608 			    ("%s: found entry pointing to a different vnode "
2609 			    "(%p != %p); name [%s]",
2610 			    __func__, n2->nc_vp, vp, cnp->cn_nameptr));
2611 		/*
2612 		 * Entries are supposed to be immutable unless in the
2613 		 * process of getting destroyed. Accommodating for
2614 		 * changing timestamps is possible but not worth it.
2615 		 * This should be harmless in terms of correctness, in
2616 		 * the worst case resulting in an earlier expiration.
2617 		 * Alternatively, the found entry can be replaced
2618 		 * altogether.
2619 		 */
2620 		MPASS((n2->nc_flag & (NCF_TS | NCF_DTS)) ==
2621 		    (ncp->nc_flag & (NCF_TS | NCF_DTS)));
2622 #if 0
2623 		if (tsp != NULL) {
2624 			KASSERT((n2->nc_flag & NCF_TS) != 0,
2625 			    ("no NCF_TS"));
2626 			n2_ts = __containerof(n2, struct namecache_ts, nc_nc);
2627 			n2_ts->nc_time = ncp_ts->nc_time;
2628 			n2_ts->nc_ticks = ncp_ts->nc_ticks;
2629 			if (dtsp != NULL) {
2630 				n2_ts->nc_dotdottime = ncp_ts->nc_dotdottime;
2631 				n2_ts->nc_nc.nc_flag |= NCF_DTS;
2632 			}
2633 		}
2634 #endif
2635 		SDT_PROBE3(vfs, namecache, enter, duplicate, dvp, ncp->nc_name,
2636 		    vp);
2637 		goto out_unlock_free;
2638 	}
2639 
2640 	if (flag == NCF_ISDOTDOT) {
2641 		/*
2642 		 * See if we are trying to add .. entry, but some other lookup
2643 		 * has populated v_cache_dd pointer already.
2644 		 */
2645 		if (dvp->v_cache_dd != NULL)
2646 			goto out_unlock_free;
2647 		KASSERT(vp == NULL || vp->v_type == VDIR,
2648 		    ("wrong vnode type %p", vp));
2649 		atomic_thread_fence_rel();
2650 		atomic_store_ptr(&dvp->v_cache_dd, ncp);
2651 	} else if (vp != NULL) {
2652 		/*
2653 		 * Take the slow path in INOTIFY().  This flag will be lazily
2654 		 * cleared by cache_vop_inotify() once all directories referring
2655 		 * to vp are unwatched.
2656 		 */
2657 		if (__predict_false((vn_irflag_read(dvp) & VIRF_INOTIFY) != 0))
2658 			vn_irflag_set_cond(vp, VIRF_INOTIFY_PARENT);
2659 
2660 		/*
2661 		 * For this case, the cache entry maps both the
2662 		 * directory name in it and the name ".." for the
2663 		 * directory's parent.
2664 		 */
2665 		if ((ndd = vp->v_cache_dd) != NULL) {
2666 			if ((ndd->nc_flag & NCF_ISDOTDOT) != 0)
2667 				cache_zap_locked(ndd);
2668 			else
2669 				ndd = NULL;
2670 		}
2671 		atomic_thread_fence_rel();
2672 		atomic_store_ptr(&vp->v_cache_dd, ncp);
2673 	}
2674 
2675 	if (flag != NCF_ISDOTDOT) {
2676 		if (LIST_EMPTY(&dvp->v_cache_src)) {
2677 			cache_hold_vnode(dvp);
2678 		}
2679 		LIST_INSERT_HEAD(&dvp->v_cache_src, ncp, nc_src);
2680 	}
2681 
2682 	/*
2683 	 * If the entry is "negative", we place it into the
2684 	 * "negative" cache queue, otherwise, we place it into the
2685 	 * destination vnode's cache entries queue.
2686 	 */
2687 	if (vp != NULL) {
2688 		TAILQ_INSERT_HEAD(&vp->v_cache_dst, ncp, nc_dst);
2689 		SDT_PROBE3(vfs, namecache, enter, done, dvp, ncp->nc_name,
2690 		    vp);
2691 	} else {
2692 		if (cnp->cn_flags & ISWHITEOUT)
2693 			atomic_store_char(&ncp->nc_flag, ncp->nc_flag | NCF_WHITE);
2694 		cache_neg_insert(ncp);
2695 		SDT_PROBE2(vfs, namecache, enter_negative, done, dvp,
2696 		    ncp->nc_name);
2697 	}
2698 
2699 	/*
2700 	 * Insert the new namecache entry into the appropriate chain
2701 	 * within the cache entries table.
2702 	 */
2703 	CK_SLIST_INSERT_HEAD(NCHHASH(hash), ncp, nc_hash);
2704 
2705 	atomic_thread_fence_rel();
2706 	/*
2707 	 * Mark the entry as fully constructed.
2708 	 * It is immutable past this point until its removal.
2709 	 */
2710 	atomic_store_char(&ncp->nc_flag, ncp->nc_flag & ~NCF_WIP);
2711 
2712 	cache_enter_unlock(&cel);
2713 	if (ndd != NULL)
2714 		cache_free(ndd);
2715 	return;
2716 out_unlock_free:
2717 	cache_enter_unlock(&cel);
2718 	cache_free(ncp);
2719 	return;
2720 }
2721 
2722 /*
2723  * A variant of the above accepting flags.
2724  *
2725  * - VFS_CACHE_DROPOLD -- if a conflicting entry is found, drop it.
2726  *
2727  * TODO: this routine is a hack. It blindly removes the old entry, even if it
2728  * happens to match and it is doing it in an inefficient manner. It was added
2729  * to accommodate NFS which runs into a case where the target for a given name
2730  * may change from under it. Note this does nothing to solve the following
2731  * race: 2 callers of cache_enter_time_flags pass a different target vnode for
2732  * the same [dvp, cnp]. It may be argued that code doing this is broken.
2733  */
2734 void
cache_enter_time_flags(struct vnode * dvp,struct vnode * vp,struct componentname * cnp,struct timespec * tsp,struct timespec * dtsp,int flags)2735 cache_enter_time_flags(struct vnode *dvp, struct vnode *vp, struct componentname *cnp,
2736     struct timespec *tsp, struct timespec *dtsp, int flags)
2737 {
2738 
2739 	MPASS((flags & ~(VFS_CACHE_DROPOLD)) == 0);
2740 
2741 	if (flags & VFS_CACHE_DROPOLD)
2742 		cache_remove_cnp(dvp, cnp);
2743 	cache_enter_time(dvp, vp, cnp, tsp, dtsp);
2744 }
2745 
2746 static u_long
cache_roundup_2(u_long val)2747 cache_roundup_2(u_long val)
2748 {
2749 	u_long res;
2750 
2751 	for (res = 1; res <= val; res <<= 1)
2752 		continue;
2753 
2754 	return (res);
2755 }
2756 
2757 static struct nchashhead *
nchinittbl(u_long elements,u_long * hashmask)2758 nchinittbl(u_long elements, u_long *hashmask)
2759 {
2760 	struct nchashhead *hashtbl;
2761 	u_long hashsize, i;
2762 
2763 	hashsize = cache_roundup_2(elements) / 2;
2764 
2765 	hashtbl = malloc(hashsize * sizeof(*hashtbl), M_VFSCACHE, M_WAITOK);
2766 	for (i = 0; i < hashsize; i++)
2767 		CK_SLIST_INIT(&hashtbl[i]);
2768 	*hashmask = hashsize - 1;
2769 	return (hashtbl);
2770 }
2771 
2772 static void
ncfreetbl(struct nchashhead * hashtbl)2773 ncfreetbl(struct nchashhead *hashtbl)
2774 {
2775 
2776 	free(hashtbl, M_VFSCACHE);
2777 }
2778 
2779 /*
2780  * Name cache initialization, from vfs_init() when we are booting
2781  */
2782 static void
nchinit(void * dummy __unused)2783 nchinit(void *dummy __unused)
2784 {
2785 	u_int i;
2786 
2787 	cache_zone_small = uma_zcreate("S VFS Cache", CACHE_ZONE_SMALL_SIZE,
2788 	    NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2789 	cache_zone_small_ts = uma_zcreate("STS VFS Cache", CACHE_ZONE_SMALL_TS_SIZE,
2790 	    NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2791 	cache_zone_large = uma_zcreate("L VFS Cache", CACHE_ZONE_LARGE_SIZE,
2792 	    NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2793 	cache_zone_large_ts = uma_zcreate("LTS VFS Cache", CACHE_ZONE_LARGE_TS_SIZE,
2794 	    NULL, NULL, NULL, NULL, CACHE_ZONE_ALIGNMENT, UMA_ZONE_ZINIT);
2795 
2796 	VFS_SMR_ZONE_SET(cache_zone_small);
2797 	VFS_SMR_ZONE_SET(cache_zone_small_ts);
2798 	VFS_SMR_ZONE_SET(cache_zone_large);
2799 	VFS_SMR_ZONE_SET(cache_zone_large_ts);
2800 
2801 	ncsize = desiredvnodes * ncsizefactor;
2802 	cache_recalc_neg_min();
2803 	nchashtbl = nchinittbl(ncsize, &nchash);
2804 	ncbuckethash = cache_roundup_2(mp_ncpus * mp_ncpus) - 1;
2805 	if (ncbuckethash < 7) /* arbitrarily chosen to avoid having one lock */
2806 		ncbuckethash = 7;
2807 	if (ncbuckethash > nchash)
2808 		ncbuckethash = nchash;
2809 	bucketlocks = malloc(sizeof(*bucketlocks) * numbucketlocks, M_VFSCACHE,
2810 	    M_WAITOK | M_ZERO);
2811 	for (i = 0; i < numbucketlocks; i++)
2812 		mtx_init(&bucketlocks[i], "ncbuc", NULL, MTX_DUPOK | MTX_RECURSE);
2813 	ncvnodehash = ncbuckethash;
2814 	vnodelocks = malloc(sizeof(*vnodelocks) * numvnodelocks, M_VFSCACHE,
2815 	    M_WAITOK | M_ZERO);
2816 	for (i = 0; i < numvnodelocks; i++)
2817 		mtx_init(&vnodelocks[i], "ncvn", NULL, MTX_DUPOK | MTX_RECURSE);
2818 
2819 	for (i = 0; i < numneglists; i++) {
2820 		mtx_init(&neglists[i].nl_evict_lock, "ncnege", NULL, MTX_DEF);
2821 		mtx_init(&neglists[i].nl_lock, "ncnegl", NULL, MTX_DEF);
2822 		TAILQ_INIT(&neglists[i].nl_list);
2823 		TAILQ_INIT(&neglists[i].nl_hotlist);
2824 	}
2825 }
2826 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_SECOND, nchinit, NULL);
2827 
2828 void
cache_vnode_init(struct vnode * vp)2829 cache_vnode_init(struct vnode *vp)
2830 {
2831 
2832 	LIST_INIT(&vp->v_cache_src);
2833 	TAILQ_INIT(&vp->v_cache_dst);
2834 	vp->v_cache_dd = NULL;
2835 	cache_prehash(vp);
2836 }
2837 
2838 /*
2839  * Induce transient cache misses for lockless operation in cache_lookup() by
2840  * using a temporary hash table.
2841  *
2842  * This will force a fs lookup.
2843  *
2844  * Synchronisation is done in 2 steps, calling vfs_smr_synchronize each time
2845  * to observe all CPUs not performing the lookup.
2846  */
2847 static void
cache_changesize_set_temp(struct nchashhead * temptbl,u_long temphash)2848 cache_changesize_set_temp(struct nchashhead *temptbl, u_long temphash)
2849 {
2850 
2851 	MPASS(temphash < nchash);
2852 	/*
2853 	 * Change the size. The new size is smaller and can safely be used
2854 	 * against the existing table. All lookups which now hash wrong will
2855 	 * result in a cache miss, which all callers are supposed to know how
2856 	 * to handle.
2857 	 */
2858 	atomic_store_long(&nchash, temphash);
2859 	atomic_thread_fence_rel();
2860 	vfs_smr_synchronize();
2861 	/*
2862 	 * At this point everyone sees the updated hash value, but they still
2863 	 * see the old table.
2864 	 */
2865 	atomic_store_ptr(&nchashtbl, temptbl);
2866 	atomic_thread_fence_rel();
2867 	vfs_smr_synchronize();
2868 	/*
2869 	 * At this point everyone sees the updated table pointer and size pair.
2870 	 */
2871 }
2872 
2873 /*
2874  * Set the new hash table.
2875  *
2876  * Similarly to cache_changesize_set_temp(), this has to synchronize against
2877  * lockless operation in cache_lookup().
2878  */
2879 static void
cache_changesize_set_new(struct nchashhead * new_tbl,u_long new_hash)2880 cache_changesize_set_new(struct nchashhead *new_tbl, u_long new_hash)
2881 {
2882 
2883 	MPASS(nchash < new_hash);
2884 	/*
2885 	 * Change the pointer first. This wont result in out of bounds access
2886 	 * since the temporary table is guaranteed to be smaller.
2887 	 */
2888 	atomic_store_ptr(&nchashtbl, new_tbl);
2889 	atomic_thread_fence_rel();
2890 	vfs_smr_synchronize();
2891 	/*
2892 	 * At this point everyone sees the updated pointer value, but they
2893 	 * still see the old size.
2894 	 */
2895 	atomic_store_long(&nchash, new_hash);
2896 	atomic_thread_fence_rel();
2897 	vfs_smr_synchronize();
2898 	/*
2899 	 * At this point everyone sees the updated table pointer and size pair.
2900 	 */
2901 }
2902 
2903 void
cache_changesize(u_long newmaxvnodes)2904 cache_changesize(u_long newmaxvnodes)
2905 {
2906 	struct nchashhead *new_nchashtbl, *old_nchashtbl, *temptbl;
2907 	u_long new_nchash, old_nchash, temphash;
2908 	struct namecache *ncp;
2909 	uint32_t hash;
2910 	u_long newncsize;
2911 	u_long i;
2912 
2913 	newncsize = newmaxvnodes * ncsizefactor;
2914 	newmaxvnodes = cache_roundup_2(newmaxvnodes * 2);
2915 	if (newmaxvnodes < numbucketlocks)
2916 		newmaxvnodes = numbucketlocks;
2917 
2918 	new_nchashtbl = nchinittbl(newmaxvnodes, &new_nchash);
2919 	/* If same hash table size, nothing to do */
2920 	if (nchash == new_nchash) {
2921 		ncfreetbl(new_nchashtbl);
2922 		return;
2923 	}
2924 
2925 	temptbl = nchinittbl(1, &temphash);
2926 
2927 	/*
2928 	 * Move everything from the old hash table to the new table.
2929 	 * None of the namecache entries in the table can be removed
2930 	 * because to do so, they have to be removed from the hash table.
2931 	 */
2932 	cache_lock_all_vnodes();
2933 	cache_lock_all_buckets();
2934 	old_nchashtbl = nchashtbl;
2935 	old_nchash = nchash;
2936 	cache_changesize_set_temp(temptbl, temphash);
2937 	for (i = 0; i <= old_nchash; i++) {
2938 		while ((ncp = CK_SLIST_FIRST(&old_nchashtbl[i])) != NULL) {
2939 			hash = cache_get_hash(ncp->nc_name, ncp->nc_nlen,
2940 			    ncp->nc_dvp);
2941 			CK_SLIST_REMOVE(&old_nchashtbl[i], ncp, namecache, nc_hash);
2942 			CK_SLIST_INSERT_HEAD(&new_nchashtbl[hash & new_nchash], ncp, nc_hash);
2943 		}
2944 	}
2945 	ncsize = newncsize;
2946 	cache_recalc_neg_min();
2947 	cache_changesize_set_new(new_nchashtbl, new_nchash);
2948 	cache_unlock_all_buckets();
2949 	cache_unlock_all_vnodes();
2950 	ncfreetbl(old_nchashtbl);
2951 	ncfreetbl(temptbl);
2952 }
2953 
2954 /*
2955  * Remove all entries from and to a particular vnode.
2956  */
2957 static void
cache_purge_impl(struct vnode * vp)2958 cache_purge_impl(struct vnode *vp)
2959 {
2960 	struct cache_freebatch batch;
2961 	struct namecache *ncp;
2962 	struct mtx *vlp, *vlp2;
2963 
2964 	TAILQ_INIT(&batch);
2965 	vlp = VP2VNODELOCK(vp);
2966 	vlp2 = NULL;
2967 	mtx_lock(vlp);
2968 retry:
2969 	while (!LIST_EMPTY(&vp->v_cache_src)) {
2970 		ncp = LIST_FIRST(&vp->v_cache_src);
2971 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2972 			goto retry;
2973 		TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2974 	}
2975 	while (!TAILQ_EMPTY(&vp->v_cache_dst)) {
2976 		ncp = TAILQ_FIRST(&vp->v_cache_dst);
2977 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2978 			goto retry;
2979 		TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2980 	}
2981 	ncp = vp->v_cache_dd;
2982 	if (ncp != NULL) {
2983 		KASSERT(ncp->nc_flag & NCF_ISDOTDOT,
2984 		   ("lost dotdot link"));
2985 		if (!cache_zap_locked_vnode_kl2(ncp, vp, &vlp2))
2986 			goto retry;
2987 		TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
2988 	}
2989 	KASSERT(vp->v_cache_dd == NULL, ("incomplete purge"));
2990 	mtx_unlock(vlp);
2991 	if (vlp2 != NULL)
2992 		mtx_unlock(vlp2);
2993 	cache_free_batch(&batch);
2994 }
2995 
2996 /*
2997  * Opportunistic check to see if there is anything to do.
2998  */
2999 static bool
cache_has_entries(struct vnode * vp)3000 cache_has_entries(struct vnode *vp)
3001 {
3002 
3003 	if (LIST_EMPTY(&vp->v_cache_src) && TAILQ_EMPTY(&vp->v_cache_dst) &&
3004 	    atomic_load_ptr(&vp->v_cache_dd) == NULL)
3005 		return (false);
3006 	return (true);
3007 }
3008 
3009 void
cache_purge(struct vnode * vp)3010 cache_purge(struct vnode *vp)
3011 {
3012 
3013 	SDT_PROBE1(vfs, namecache, purge, done, vp);
3014 	if (!cache_has_entries(vp))
3015 		return;
3016 	cache_purge_impl(vp);
3017 }
3018 
3019 /*
3020  * Only to be used by vgone.
3021  */
3022 void
cache_purge_vgone(struct vnode * vp)3023 cache_purge_vgone(struct vnode *vp)
3024 {
3025 	struct mtx *vlp;
3026 
3027 	VNPASS(VN_IS_DOOMED(vp), vp);
3028 	if (cache_has_entries(vp)) {
3029 		cache_purge_impl(vp);
3030 		return;
3031 	}
3032 
3033 	/*
3034 	 * Serialize against a potential thread doing cache_purge.
3035 	 */
3036 	vlp = VP2VNODELOCK(vp);
3037 	mtx_wait_unlocked(vlp);
3038 	if (cache_has_entries(vp)) {
3039 		cache_purge_impl(vp);
3040 		return;
3041 	}
3042 	return;
3043 }
3044 
3045 /*
3046  * Remove all negative entries for a particular directory vnode.
3047  */
3048 void
cache_purge_negative(struct vnode * vp)3049 cache_purge_negative(struct vnode *vp)
3050 {
3051 	struct cache_freebatch batch;
3052 	struct namecache *ncp, *nnp;
3053 	struct mtx *vlp;
3054 
3055 	SDT_PROBE1(vfs, namecache, purge_negative, done, vp);
3056 	if (LIST_EMPTY(&vp->v_cache_src))
3057 		return;
3058 	TAILQ_INIT(&batch);
3059 	vlp = VP2VNODELOCK(vp);
3060 	mtx_lock(vlp);
3061 	LIST_FOREACH_SAFE(ncp, &vp->v_cache_src, nc_src, nnp) {
3062 		if (!(ncp->nc_flag & NCF_NEGATIVE))
3063 			continue;
3064 		cache_zap_negative_locked_vnode_kl(ncp, vp);
3065 		TAILQ_INSERT_TAIL(&batch, ncp, nc_dst);
3066 	}
3067 	mtx_unlock(vlp);
3068 	cache_free_batch(&batch);
3069 }
3070 
3071 /*
3072  * Entry points for modifying VOP operations.
3073  */
3074 void
cache_vop_rename(struct vnode * fdvp,struct vnode * fvp,struct vnode * tdvp,struct vnode * tvp,struct componentname * fcnp,struct componentname * tcnp)3075 cache_vop_rename(struct vnode *fdvp, struct vnode *fvp, struct vnode *tdvp,
3076     struct vnode *tvp, struct componentname *fcnp, struct componentname *tcnp)
3077 {
3078 
3079 	ASSERT_VOP_IN_SEQC(fdvp);
3080 	ASSERT_VOP_IN_SEQC(fvp);
3081 	ASSERT_VOP_IN_SEQC(tdvp);
3082 	if (tvp != NULL)
3083 		ASSERT_VOP_IN_SEQC(tvp);
3084 
3085 	cache_purge(fvp);
3086 	if (tvp != NULL) {
3087 		cache_purge(tvp);
3088 		KASSERT(!cache_remove_cnp(tdvp, tcnp),
3089 		    ("%s: lingering negative entry", __func__));
3090 	} else {
3091 		cache_remove_cnp(tdvp, tcnp);
3092 	}
3093 
3094 	/*
3095 	 * TODO
3096 	 *
3097 	 * Historically renaming was always purging all revelang entries,
3098 	 * but that's quite wasteful. In particular turns out that in many cases
3099 	 * the target file is immediately accessed after rename, inducing a cache
3100 	 * miss.
3101 	 *
3102 	 * Recode this to reduce relocking and reuse the existing entry (if any)
3103 	 * instead of just removing it above and allocating a new one here.
3104 	 */
3105 	cache_enter(tdvp, fvp, tcnp);
3106 }
3107 
3108 void
cache_vop_rmdir(struct vnode * dvp,struct vnode * vp)3109 cache_vop_rmdir(struct vnode *dvp, struct vnode *vp)
3110 {
3111 
3112 	ASSERT_VOP_IN_SEQC(dvp);
3113 	ASSERT_VOP_IN_SEQC(vp);
3114 	cache_purge(vp);
3115 }
3116 
3117 #ifdef INVARIANTS
3118 /*
3119  * Validate that if an entry exists it matches.
3120  */
3121 void
cache_validate(struct vnode * dvp,struct vnode * vp,struct componentname * cnp)3122 cache_validate(struct vnode *dvp, struct vnode *vp, struct componentname *cnp)
3123 {
3124 	struct namecache *ncp;
3125 	struct mtx *blp;
3126 	uint32_t hash;
3127 
3128 	hash = cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp);
3129 	if (CK_SLIST_EMPTY(NCHHASH(hash)))
3130 		return;
3131 	blp = HASH2BUCKETLOCK(hash);
3132 	mtx_lock(blp);
3133 	ncp = cache_ncp_find(dvp, cnp, hash);
3134 	if (ncp != NULL && ncp->nc_vp != vp) {
3135 		panic("%s: mismatch (%p != %p); ncp %p [%s] dvp %p\n",
3136 		    __func__, vp, ncp->nc_vp, ncp, ncp->nc_name, ncp->nc_dvp);
3137 	}
3138 	mtx_unlock(blp);
3139 }
3140 
3141 void
cache_assert_no_entries(struct vnode * vp)3142 cache_assert_no_entries(struct vnode *vp)
3143 {
3144 
3145 	VNPASS(TAILQ_EMPTY(&vp->v_cache_dst), vp);
3146 	VNPASS(LIST_EMPTY(&vp->v_cache_src), vp);
3147 	VNPASS(vp->v_cache_dd == NULL, vp);
3148 }
3149 #endif
3150 
3151 /*
3152  * Flush all entries referencing a particular filesystem.
3153  */
3154 void
cache_purgevfs(struct mount * mp)3155 cache_purgevfs(struct mount *mp)
3156 {
3157 	struct vnode *vp, *mvp;
3158 	size_t visited __sdt_used, purged __sdt_used;
3159 
3160 	visited = purged = 0;
3161 	/*
3162 	 * Somewhat wasteful iteration over all vnodes. Would be better to
3163 	 * support filtering and avoid the interlock to begin with.
3164 	 */
3165 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3166 		visited++;
3167 		if (!cache_has_entries(vp)) {
3168 			VI_UNLOCK(vp);
3169 			continue;
3170 		}
3171 		vholdl(vp);
3172 		VI_UNLOCK(vp);
3173 		cache_purge(vp);
3174 		purged++;
3175 		vdrop(vp);
3176 	}
3177 
3178 	SDT_PROBE3(vfs, namecache, purgevfs, done, mp, visited, purged);
3179 }
3180 
3181 /*
3182  * Perform canonical checks and cache lookup and pass on to filesystem
3183  * through the vop_cachedlookup only if needed.
3184  */
3185 
3186 int
vfs_cache_lookup(struct vop_lookup_args * ap)3187 vfs_cache_lookup(struct vop_lookup_args *ap)
3188 {
3189 	struct vnode *dvp;
3190 	int error;
3191 	struct vnode **vpp = ap->a_vpp;
3192 	struct componentname *cnp = ap->a_cnp;
3193 	int flags = cnp->cn_flags;
3194 
3195 	*vpp = NULL;
3196 	dvp = ap->a_dvp;
3197 
3198 	if (dvp->v_type != VDIR)
3199 		return (ENOTDIR);
3200 
3201 	if ((flags & ISLASTCN) && (dvp->v_mount->mnt_flag & MNT_RDONLY) &&
3202 	    (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME))
3203 		return (EROFS);
3204 
3205 	error = vn_dir_check_exec(dvp, cnp);
3206 	if (error != 0)
3207 		return (error);
3208 
3209 	error = cache_lookup(dvp, vpp, cnp, NULL, NULL);
3210 	if (error == 0)
3211 		return (VOP_CACHEDLOOKUP(dvp, vpp, cnp));
3212 	if (error == -1)
3213 		return (0);
3214 	return (error);
3215 }
3216 
3217 /* Implementation of the getcwd syscall. */
3218 int
sys___getcwd(struct thread * td,struct __getcwd_args * uap)3219 sys___getcwd(struct thread *td, struct __getcwd_args *uap)
3220 {
3221 	char *buf, *retbuf;
3222 	size_t buflen;
3223 	int error;
3224 
3225 	buflen = uap->buflen;
3226 	if (__predict_false(buflen < 2))
3227 		return (EINVAL);
3228 	if (buflen > MAXPATHLEN)
3229 		buflen = MAXPATHLEN;
3230 
3231 	buf = uma_zalloc(namei_zone, M_WAITOK);
3232 	error = vn_getcwd(buf, &retbuf, &buflen);
3233 	if (error == 0)
3234 		error = copyout(retbuf, uap->buf, buflen);
3235 	uma_zfree(namei_zone, buf);
3236 	return (error);
3237 }
3238 
3239 int
vn_getcwd(char * buf,char ** retbuf,size_t * buflen)3240 vn_getcwd(char *buf, char **retbuf, size_t *buflen)
3241 {
3242 	struct pwd *pwd;
3243 	int error;
3244 
3245 	vfs_smr_enter();
3246 	pwd = pwd_get_smr();
3247 	error = vn_fullpath_any_smr(pwd->pwd_cdir, pwd->pwd_rdir, buf, retbuf,
3248 	    buflen, 0);
3249 	VFS_SMR_ASSERT_NOT_ENTERED();
3250 	if (error < 0) {
3251 		pwd = pwd_hold(curthread);
3252 		error = vn_fullpath_any(pwd->pwd_cdir, pwd->pwd_rdir, buf,
3253 		    retbuf, buflen);
3254 		pwd_drop(pwd);
3255 	}
3256 
3257 #ifdef KTRACE
3258 	if (KTRPOINT(curthread, KTR_NAMEI) && error == 0)
3259 		ktrnamei(*retbuf);
3260 #endif
3261 	return (error);
3262 }
3263 
3264 /*
3265  * Canonicalize a path by walking it forward and back.
3266  *
3267  * BUGS:
3268  * - Nothing guarantees the integrity of the entire chain. Consider the case
3269  *   where the path "foo/bar/baz/qux" is passed, but "bar" is moved out of
3270  *   "foo" into "quux" during the backwards walk. The result will be
3271  *   "quux/bar/baz/qux", which could not have been obtained by an incremental
3272  *   walk in userspace. Moreover, the path we return is inaccessible if the
3273  *   calling thread lacks permission to traverse "quux".
3274  */
3275 static int
kern___realpathat(struct thread * td,int fd,const char * path,char * buf,size_t size,int flags,enum uio_seg pathseg)3276 kern___realpathat(struct thread *td, int fd, const char *path, char *buf,
3277     size_t size, int flags, enum uio_seg pathseg)
3278 {
3279 	struct nameidata nd;
3280 	char *retbuf, *freebuf;
3281 	int error;
3282 
3283 	if (flags != 0)
3284 		return (EINVAL);
3285 	NDINIT_ATRIGHTS(&nd, LOOKUP, FOLLOW | WANTPARENT | AUDITVNODE1,
3286 	    pathseg, path, fd, &cap_fstat_rights);
3287 	if ((error = namei(&nd)) != 0)
3288 		return (error);
3289 
3290 	if (nd.ni_vp->v_type == VREG && nd.ni_dvp->v_type != VDIR &&
3291 	    (nd.ni_vp->v_vflag & VV_ROOT) != 0) {
3292 		struct vnode *covered_vp;
3293 
3294 		/*
3295 		 * This happens if vp is a file mount. The call to
3296 		 * vn_fullpath_hardlink can panic if path resolution can't be
3297 		 * handled without the directory.
3298 		 *
3299 		 * To resolve this, we find the vnode which was mounted on -
3300 		 * this should have a unique global path since we disallow
3301 		 * mounting on linked files.
3302 		 */
3303 		error = vn_lock(nd.ni_vp, LK_SHARED);
3304 		if (error != 0)
3305 			goto out;
3306 		covered_vp = nd.ni_vp->v_mount->mnt_vnodecovered;
3307 		vref(covered_vp);
3308 		VOP_UNLOCK(nd.ni_vp);
3309 		error = vn_fullpath(covered_vp, &retbuf, &freebuf);
3310 		vrele(covered_vp);
3311 	} else {
3312 		error = vn_fullpath_hardlink(nd.ni_vp, nd.ni_dvp,
3313 		    nd.ni_cnd.cn_nameptr, nd.ni_cnd.cn_namelen, &retbuf,
3314 		    &freebuf, &size);
3315 	}
3316 	if (error == 0) {
3317 		size_t len;
3318 
3319 		len = strlen(retbuf) + 1;
3320 		if (size < len)
3321 			error = ENAMETOOLONG;
3322 		else if (pathseg == UIO_USERSPACE)
3323 			error = copyout(retbuf, buf, len);
3324 		else
3325 			memcpy(buf, retbuf, len);
3326 		free(freebuf, M_TEMP);
3327 	}
3328 out:
3329 	vrele(nd.ni_vp);
3330 	vrele(nd.ni_dvp);
3331 	NDFREE_PNBUF(&nd);
3332 	return (error);
3333 }
3334 
3335 int
sys___realpathat(struct thread * td,struct __realpathat_args * uap)3336 sys___realpathat(struct thread *td, struct __realpathat_args *uap)
3337 {
3338 
3339 	return (kern___realpathat(td, uap->fd, uap->path, uap->buf, uap->size,
3340 	    uap->flags, UIO_USERSPACE));
3341 }
3342 
3343 /*
3344  * Retrieve the full filesystem path that correspond to a vnode from the name
3345  * cache (if available)
3346  */
3347 int
vn_fullpath(struct vnode * vp,char ** retbuf,char ** freebuf)3348 vn_fullpath(struct vnode *vp, char **retbuf, char **freebuf)
3349 {
3350 	struct pwd *pwd;
3351 	char *buf;
3352 	size_t buflen;
3353 	int error;
3354 
3355 	if (__predict_false(vp == NULL))
3356 		return (EINVAL);
3357 
3358 	buflen = MAXPATHLEN;
3359 	buf = malloc(buflen, M_TEMP, M_WAITOK);
3360 	vfs_smr_enter();
3361 	pwd = pwd_get_smr();
3362 	error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, &buflen, 0);
3363 	VFS_SMR_ASSERT_NOT_ENTERED();
3364 	if (error < 0) {
3365 		pwd = pwd_hold(curthread);
3366 		error = vn_fullpath_any(vp, pwd->pwd_rdir, buf, retbuf, &buflen);
3367 		pwd_drop(pwd);
3368 	}
3369 	if (error == 0)
3370 		*freebuf = buf;
3371 	else
3372 		free(buf, M_TEMP);
3373 	return (error);
3374 }
3375 
3376 /*
3377  * This function is similar to vn_fullpath, but it attempts to lookup the
3378  * pathname relative to the global root mount point.  This is required for the
3379  * auditing sub-system, as audited pathnames must be absolute, relative to the
3380  * global root mount point.
3381  */
3382 int
vn_fullpath_global(struct vnode * vp,char ** retbuf,char ** freebuf)3383 vn_fullpath_global(struct vnode *vp, char **retbuf, char **freebuf)
3384 {
3385 	char *buf;
3386 	size_t buflen;
3387 	int error;
3388 
3389 	if (__predict_false(vp == NULL))
3390 		return (EINVAL);
3391 	buflen = MAXPATHLEN;
3392 	buf = malloc(buflen, M_TEMP, M_WAITOK);
3393 	vfs_smr_enter();
3394 	error = vn_fullpath_any_smr(vp, rootvnode, buf, retbuf, &buflen, 0);
3395 	VFS_SMR_ASSERT_NOT_ENTERED();
3396 	if (error < 0) {
3397 		error = vn_fullpath_any(vp, rootvnode, buf, retbuf, &buflen);
3398 	}
3399 	if (error == 0)
3400 		*freebuf = buf;
3401 	else
3402 		free(buf, M_TEMP);
3403 	return (error);
3404 }
3405 
3406 static struct namecache *
vn_dd_from_dst(struct vnode * vp)3407 vn_dd_from_dst(struct vnode *vp)
3408 {
3409 	struct namecache *ncp;
3410 
3411 	cache_assert_vnode_locked(vp);
3412 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) {
3413 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
3414 			return (ncp);
3415 	}
3416 	return (NULL);
3417 }
3418 
3419 int
vn_vptocnp(struct vnode ** vp,char * buf,size_t * buflen)3420 vn_vptocnp(struct vnode **vp, char *buf, size_t *buflen)
3421 {
3422 	struct vnode *dvp;
3423 	struct namecache *ncp;
3424 	struct mtx *vlp;
3425 	int error;
3426 
3427 	vlp = VP2VNODELOCK(*vp);
3428 	mtx_lock(vlp);
3429 	ncp = (*vp)->v_cache_dd;
3430 	if (ncp != NULL && (ncp->nc_flag & NCF_ISDOTDOT) == 0) {
3431 		KASSERT(ncp == vn_dd_from_dst(*vp),
3432 		    ("%s: mismatch for dd entry (%p != %p)", __func__,
3433 		    ncp, vn_dd_from_dst(*vp)));
3434 	} else {
3435 		ncp = vn_dd_from_dst(*vp);
3436 	}
3437 	if (ncp != NULL) {
3438 		if (*buflen < ncp->nc_nlen) {
3439 			mtx_unlock(vlp);
3440 			vrele(*vp);
3441 			counter_u64_add(numfullpathfail4, 1);
3442 			error = ENOMEM;
3443 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
3444 			    vp, NULL);
3445 			return (error);
3446 		}
3447 		*buflen -= ncp->nc_nlen;
3448 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
3449 		SDT_PROBE3(vfs, namecache, fullpath, hit, ncp->nc_dvp,
3450 		    ncp->nc_name, vp);
3451 		dvp = *vp;
3452 		*vp = ncp->nc_dvp;
3453 		vref(*vp);
3454 		mtx_unlock(vlp);
3455 		vrele(dvp);
3456 		return (0);
3457 	}
3458 	SDT_PROBE1(vfs, namecache, fullpath, miss, vp);
3459 
3460 	mtx_unlock(vlp);
3461 	vn_lock(*vp, LK_SHARED | LK_RETRY);
3462 	error = VOP_VPTOCNP(*vp, &dvp, buf, buflen);
3463 	vput(*vp);
3464 	if (error) {
3465 		counter_u64_add(numfullpathfail2, 1);
3466 		SDT_PROBE3(vfs, namecache, fullpath, return,  error, vp, NULL);
3467 		return (error);
3468 	}
3469 
3470 	*vp = dvp;
3471 	if (VN_IS_DOOMED(dvp)) {
3472 		/* forced unmount */
3473 		vrele(dvp);
3474 		error = ENOENT;
3475 		SDT_PROBE3(vfs, namecache, fullpath, return, error, vp, NULL);
3476 		return (error);
3477 	}
3478 	/*
3479 	 * *vp has its use count incremented still.
3480 	 */
3481 
3482 	return (0);
3483 }
3484 
3485 /*
3486  * Resolve a directory to a pathname.
3487  *
3488  * The name of the directory can always be found in the namecache or fetched
3489  * from the filesystem. There is also guaranteed to be only one parent, meaning
3490  * we can just follow vnodes up until we find the root.
3491  *
3492  * The vnode must be referenced.
3493  */
3494 static int
vn_fullpath_dir(struct vnode * vp,struct vnode * rdir,char * buf,char ** retbuf,size_t * len,size_t addend)3495 vn_fullpath_dir(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3496     size_t *len, size_t addend)
3497 {
3498 #ifdef KDTRACE_HOOKS
3499 	struct vnode *startvp = vp;
3500 #endif
3501 	struct vnode *vp1;
3502 	size_t buflen;
3503 	int error;
3504 	bool slash_prefixed;
3505 
3506 	VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp);
3507 	VNPASS(vp->v_usecount > 0, vp);
3508 
3509 	buflen = *len;
3510 
3511 	slash_prefixed = true;
3512 	if (addend == 0) {
3513 		MPASS(*len >= 2);
3514 		buflen--;
3515 		buf[buflen] = '\0';
3516 		slash_prefixed = false;
3517 	}
3518 
3519 	error = 0;
3520 
3521 	SDT_PROBE1(vfs, namecache, fullpath, entry, vp);
3522 	counter_u64_add(numfullpathcalls, 1);
3523 	while (vp != rdir && vp != rootvnode) {
3524 		/*
3525 		 * The vp vnode must be already fully constructed,
3526 		 * since it is either found in namecache or obtained
3527 		 * from VOP_VPTOCNP().  We may test for VV_ROOT safely
3528 		 * without obtaining the vnode lock.
3529 		 */
3530 		if ((vp->v_vflag & VV_ROOT) != 0) {
3531 			vn_lock(vp, LK_RETRY | LK_SHARED);
3532 
3533 			/*
3534 			 * With the vnode locked, check for races with
3535 			 * unmount, forced or not.  Note that we
3536 			 * already verified that vp is not equal to
3537 			 * the root vnode, which means that
3538 			 * mnt_vnodecovered can be NULL only for the
3539 			 * case of unmount.
3540 			 */
3541 			if (VN_IS_DOOMED(vp) ||
3542 			    (vp1 = vp->v_mount->mnt_vnodecovered) == NULL ||
3543 			    vp1->v_mountedhere != vp->v_mount) {
3544 				vput(vp);
3545 				error = ENOENT;
3546 				SDT_PROBE3(vfs, namecache, fullpath, return,
3547 				    error, vp, NULL);
3548 				break;
3549 			}
3550 
3551 			vref(vp1);
3552 			vput(vp);
3553 			vp = vp1;
3554 			continue;
3555 		}
3556 		VNPASS(vp->v_type == VDIR || VN_IS_DOOMED(vp), vp);
3557 		error = vn_vptocnp(&vp, buf, &buflen);
3558 		if (error)
3559 			break;
3560 		if (buflen == 0) {
3561 			vrele(vp);
3562 			error = ENOMEM;
3563 			SDT_PROBE3(vfs, namecache, fullpath, return, error,
3564 			    startvp, NULL);
3565 			break;
3566 		}
3567 		buf[--buflen] = '/';
3568 		slash_prefixed = true;
3569 	}
3570 	if (error)
3571 		return (error);
3572 	if (!slash_prefixed) {
3573 		if (buflen == 0) {
3574 			vrele(vp);
3575 			counter_u64_add(numfullpathfail4, 1);
3576 			SDT_PROBE3(vfs, namecache, fullpath, return, ENOMEM,
3577 			    startvp, NULL);
3578 			return (ENOMEM);
3579 		}
3580 		buf[--buflen] = '/';
3581 	}
3582 	counter_u64_add(numfullpathfound, 1);
3583 	vrele(vp);
3584 
3585 	*retbuf = buf + buflen;
3586 	SDT_PROBE3(vfs, namecache, fullpath, return, 0, startvp, *retbuf);
3587 	*len -= buflen;
3588 	*len += addend;
3589 	return (0);
3590 }
3591 
3592 /*
3593  * Resolve an arbitrary vnode to a pathname.
3594  *
3595  * Note 2 caveats:
3596  * - hardlinks are not tracked, thus if the vnode is not a directory this can
3597  *   resolve to a different path than the one used to find it
3598  * - namecache is not mandatory, meaning names are not guaranteed to be added
3599  *   (in which case resolving fails)
3600  */
3601 static void __inline
cache_rev_failed_impl(int * reason,int line)3602 cache_rev_failed_impl(int *reason, int line)
3603 {
3604 
3605 	*reason = line;
3606 }
3607 #define cache_rev_failed(var)	cache_rev_failed_impl((var), __LINE__)
3608 
3609 static int
vn_fullpath_any_smr(struct vnode * vp,struct vnode * rdir,char * buf,char ** retbuf,size_t * buflen,size_t addend)3610 vn_fullpath_any_smr(struct vnode *vp, struct vnode *rdir, char *buf,
3611     char **retbuf, size_t *buflen, size_t addend)
3612 {
3613 #ifdef KDTRACE_HOOKS
3614 	struct vnode *startvp = vp;
3615 #endif
3616 	struct vnode *tvp;
3617 	struct mount *mp;
3618 	struct namecache *ncp;
3619 	size_t orig_buflen;
3620 	int reason;
3621 	int error;
3622 #ifdef KDTRACE_HOOKS
3623 	int i;
3624 #endif
3625 	seqc_t vp_seqc, tvp_seqc;
3626 	u_char nc_flag;
3627 
3628 	VFS_SMR_ASSERT_ENTERED();
3629 
3630 	if (!atomic_load_char(&cache_fast_lookup_enabled)) {
3631 		vfs_smr_exit();
3632 		return (-1);
3633 	}
3634 
3635 	orig_buflen = *buflen;
3636 
3637 	if (addend == 0) {
3638 		MPASS(*buflen >= 2);
3639 		*buflen -= 1;
3640 		buf[*buflen] = '\0';
3641 	}
3642 
3643 	if (vp == rdir || vp == rootvnode) {
3644 		if (addend == 0) {
3645 			*buflen -= 1;
3646 			buf[*buflen] = '/';
3647 		}
3648 		goto out_ok;
3649 	}
3650 
3651 #ifdef KDTRACE_HOOKS
3652 	i = 0;
3653 #endif
3654 	error = -1;
3655 	ncp = NULL; /* for sdt probe down below */
3656 	vp_seqc = vn_seqc_read_any(vp);
3657 	if (seqc_in_modify(vp_seqc)) {
3658 		cache_rev_failed(&reason);
3659 		goto out_abort;
3660 	}
3661 
3662 	for (;;) {
3663 #ifdef KDTRACE_HOOKS
3664 		i++;
3665 #endif
3666 		if ((vp->v_vflag & VV_ROOT) != 0) {
3667 			mp = atomic_load_ptr(&vp->v_mount);
3668 			if (mp == NULL) {
3669 				cache_rev_failed(&reason);
3670 				goto out_abort;
3671 			}
3672 			tvp = atomic_load_ptr(&mp->mnt_vnodecovered);
3673 			tvp_seqc = vn_seqc_read_any(tvp);
3674 			if (seqc_in_modify(tvp_seqc)) {
3675 				cache_rev_failed(&reason);
3676 				goto out_abort;
3677 			}
3678 			if (!vn_seqc_consistent(vp, vp_seqc)) {
3679 				cache_rev_failed(&reason);
3680 				goto out_abort;
3681 			}
3682 			vp = tvp;
3683 			vp_seqc = tvp_seqc;
3684 			continue;
3685 		}
3686 		ncp = atomic_load_consume_ptr(&vp->v_cache_dd);
3687 		if (ncp == NULL) {
3688 			cache_rev_failed(&reason);
3689 			goto out_abort;
3690 		}
3691 		nc_flag = atomic_load_char(&ncp->nc_flag);
3692 		if ((nc_flag & NCF_ISDOTDOT) != 0) {
3693 			cache_rev_failed(&reason);
3694 			goto out_abort;
3695 		}
3696 		if (ncp->nc_nlen >= *buflen) {
3697 			cache_rev_failed(&reason);
3698 			error = ENOMEM;
3699 			goto out_abort;
3700 		}
3701 		*buflen -= ncp->nc_nlen;
3702 		memcpy(buf + *buflen, ncp->nc_name, ncp->nc_nlen);
3703 		*buflen -= 1;
3704 		buf[*buflen] = '/';
3705 		tvp = ncp->nc_dvp;
3706 		tvp_seqc = vn_seqc_read_any(tvp);
3707 		if (seqc_in_modify(tvp_seqc)) {
3708 			cache_rev_failed(&reason);
3709 			goto out_abort;
3710 		}
3711 		if (!vn_seqc_consistent(vp, vp_seqc)) {
3712 			cache_rev_failed(&reason);
3713 			goto out_abort;
3714 		}
3715 		/*
3716 		 * Acquire fence provided by vn_seqc_read_any above.
3717 		 */
3718 		if (__predict_false(atomic_load_ptr(&vp->v_cache_dd) != ncp)) {
3719 			cache_rev_failed(&reason);
3720 			goto out_abort;
3721 		}
3722 		if (!cache_ncp_canuse(ncp)) {
3723 			cache_rev_failed(&reason);
3724 			goto out_abort;
3725 		}
3726 		vp = tvp;
3727 		vp_seqc = tvp_seqc;
3728 		if (vp == rdir || vp == rootvnode)
3729 			break;
3730 	}
3731 out_ok:
3732 	vfs_smr_exit();
3733 	*retbuf = buf + *buflen;
3734 	*buflen = orig_buflen - *buflen + addend;
3735 	SDT_PROBE2(vfs, namecache, fullpath_smr, hit, startvp, *retbuf);
3736 	return (0);
3737 
3738 out_abort:
3739 	*buflen = orig_buflen;
3740 	SDT_PROBE4(vfs, namecache, fullpath_smr, miss, startvp, ncp, reason, i);
3741 	vfs_smr_exit();
3742 	return (error);
3743 }
3744 
3745 static int
vn_fullpath_any(struct vnode * vp,struct vnode * rdir,char * buf,char ** retbuf,size_t * buflen)3746 vn_fullpath_any(struct vnode *vp, struct vnode *rdir, char *buf, char **retbuf,
3747     size_t *buflen)
3748 {
3749 	size_t orig_buflen, addend;
3750 	int error;
3751 
3752 	if (*buflen < 2)
3753 		return (EINVAL);
3754 
3755 	orig_buflen = *buflen;
3756 
3757 	vref(vp);
3758 	addend = 0;
3759 	if (vp->v_type != VDIR) {
3760 		*buflen -= 1;
3761 		buf[*buflen] = '\0';
3762 		error = vn_vptocnp(&vp, buf, buflen);
3763 		if (error)
3764 			return (error);
3765 		if (*buflen == 0) {
3766 			vrele(vp);
3767 			return (ENOMEM);
3768 		}
3769 		*buflen -= 1;
3770 		buf[*buflen] = '/';
3771 		addend = orig_buflen - *buflen;
3772 	}
3773 
3774 	return (vn_fullpath_dir(vp, rdir, buf, retbuf, buflen, addend));
3775 }
3776 
3777 /*
3778  * Resolve an arbitrary vnode to a pathname (taking care of hardlinks).
3779  *
3780  * Since the namecache does not track hardlinks, the caller is expected to
3781  * first look up the target vnode with WANTPARENT flag passed to namei to get
3782  * dvp and vp.
3783  *
3784  * Then we have 2 cases:
3785  * - if the found vnode is a directory, the path can be constructed just by
3786  *   following names up the chain
3787  * - otherwise we populate the buffer with the saved name and start resolving
3788  *   from the parent
3789  */
3790 int
vn_fullpath_hardlink(struct vnode * vp,struct vnode * dvp,const char * hrdl_name,size_t hrdl_name_length,char ** retbuf,char ** freebuf,size_t * buflen)3791 vn_fullpath_hardlink(struct vnode *vp, struct vnode *dvp,
3792     const char *hrdl_name, size_t hrdl_name_length,
3793     char **retbuf, char **freebuf, size_t *buflen)
3794 {
3795 	char *buf, *tmpbuf;
3796 	struct pwd *pwd;
3797 	size_t addend;
3798 	int error;
3799 	__enum_uint8(vtype) type;
3800 
3801 	if (*buflen < 2)
3802 		return (EINVAL);
3803 	if (*buflen > MAXPATHLEN)
3804 		*buflen = MAXPATHLEN;
3805 
3806 	buf = malloc(*buflen, M_TEMP, M_WAITOK);
3807 
3808 	addend = 0;
3809 
3810 	/*
3811 	 * Check for VBAD to work around the vp_crossmp bug in lookup().
3812 	 *
3813 	 * For example consider tmpfs on /tmp and realpath /tmp. ni_vp will be
3814 	 * set to mount point's root vnode while ni_dvp will be vp_crossmp.
3815 	 * If the type is VDIR (like in this very case) we can skip looking
3816 	 * at ni_dvp in the first place. However, since vnodes get passed here
3817 	 * unlocked the target may transition to doomed state (type == VBAD)
3818 	 * before we get to evaluate the condition. If this happens, we will
3819 	 * populate part of the buffer and descend to vn_fullpath_dir with
3820 	 * vp == vp_crossmp. Prevent the problem by checking for VBAD.
3821 	 */
3822 	type = atomic_load_8(&vp->v_type);
3823 	if (type == VBAD) {
3824 		error = ENOENT;
3825 		goto out_bad;
3826 	}
3827 	if (type != VDIR) {
3828 		addend = hrdl_name_length + 2;
3829 		if (*buflen < addend) {
3830 			error = ENOMEM;
3831 			goto out_bad;
3832 		}
3833 		*buflen -= addend;
3834 		tmpbuf = buf + *buflen;
3835 		tmpbuf[0] = '/';
3836 		memcpy(&tmpbuf[1], hrdl_name, hrdl_name_length);
3837 		tmpbuf[addend - 1] = '\0';
3838 		vp = dvp;
3839 	}
3840 
3841 	vfs_smr_enter();
3842 	pwd = pwd_get_smr();
3843 	error = vn_fullpath_any_smr(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3844 	    addend);
3845 	VFS_SMR_ASSERT_NOT_ENTERED();
3846 	if (error < 0) {
3847 		pwd = pwd_hold(curthread);
3848 		vref(vp);
3849 		error = vn_fullpath_dir(vp, pwd->pwd_rdir, buf, retbuf, buflen,
3850 		    addend);
3851 		pwd_drop(pwd);
3852 	}
3853 	if (error != 0)
3854 		goto out_bad;
3855 
3856 	*freebuf = buf;
3857 
3858 	return (0);
3859 out_bad:
3860 	free(buf, M_TEMP);
3861 	return (error);
3862 }
3863 
3864 struct vnode *
vn_dir_dd_ino(struct vnode * vp)3865 vn_dir_dd_ino(struct vnode *vp)
3866 {
3867 	struct namecache *ncp;
3868 	struct vnode *ddvp;
3869 	struct mtx *vlp;
3870 	enum vgetstate vs;
3871 
3872 	ASSERT_VOP_LOCKED(vp, "vn_dir_dd_ino");
3873 	vlp = VP2VNODELOCK(vp);
3874 	mtx_lock(vlp);
3875 	TAILQ_FOREACH(ncp, &(vp->v_cache_dst), nc_dst) {
3876 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
3877 			continue;
3878 		ddvp = ncp->nc_dvp;
3879 		vs = vget_prep(ddvp);
3880 		mtx_unlock(vlp);
3881 		if (vget_finish(ddvp, LK_SHARED | LK_NOWAIT, vs))
3882 			return (NULL);
3883 		return (ddvp);
3884 	}
3885 	mtx_unlock(vlp);
3886 	return (NULL);
3887 }
3888 
3889 int
vn_commname(struct vnode * vp,char * buf,u_int buflen)3890 vn_commname(struct vnode *vp, char *buf, u_int buflen)
3891 {
3892 	struct namecache *ncp;
3893 	struct mtx *vlp;
3894 	int l;
3895 
3896 	vlp = VP2VNODELOCK(vp);
3897 	mtx_lock(vlp);
3898 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst)
3899 		if ((ncp->nc_flag & NCF_ISDOTDOT) == 0)
3900 			break;
3901 	if (ncp == NULL) {
3902 		mtx_unlock(vlp);
3903 		return (ENOENT);
3904 	}
3905 	l = min(ncp->nc_nlen, buflen - 1);
3906 	memcpy(buf, ncp->nc_name, l);
3907 	mtx_unlock(vlp);
3908 	buf[l] = '\0';
3909 	return (0);
3910 }
3911 
3912 /*
3913  * This function updates path string to vnode's full global path
3914  * and checks the size of the new path string against the pathlen argument.
3915  *
3916  * Requires a locked, referenced vnode.
3917  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
3918  *
3919  * If vp is a directory, the call to vn_fullpath_global() always succeeds
3920  * because it falls back to the ".." lookup if the namecache lookup fails.
3921  */
3922 int
vn_path_to_global_path(struct thread * td,struct vnode * vp,char * path,u_int pathlen)3923 vn_path_to_global_path(struct thread *td, struct vnode *vp, char *path,
3924     u_int pathlen)
3925 {
3926 	struct nameidata nd;
3927 	struct vnode *vp1;
3928 	char *rpath, *fbuf;
3929 	int error;
3930 
3931 	ASSERT_VOP_ELOCKED(vp, __func__);
3932 
3933 	/* Construct global filesystem path from vp. */
3934 	VOP_UNLOCK(vp);
3935 	error = vn_fullpath_global(vp, &rpath, &fbuf);
3936 
3937 	if (error != 0) {
3938 		vrele(vp);
3939 		return (error);
3940 	}
3941 
3942 	if (strlen(rpath) >= pathlen) {
3943 		vrele(vp);
3944 		error = ENAMETOOLONG;
3945 		goto out;
3946 	}
3947 
3948 	/*
3949 	 * Re-lookup the vnode by path to detect a possible rename.
3950 	 * As a side effect, the vnode is relocked.
3951 	 * If vnode was renamed, return ENOENT.
3952 	 */
3953 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1, UIO_SYSSPACE, path);
3954 	error = namei(&nd);
3955 	if (error != 0) {
3956 		vrele(vp);
3957 		goto out;
3958 	}
3959 	NDFREE_PNBUF(&nd);
3960 	vp1 = nd.ni_vp;
3961 	vrele(vp);
3962 	if (vp1 == vp)
3963 		strcpy(path, rpath);
3964 	else {
3965 		vput(vp1);
3966 		error = ENOENT;
3967 	}
3968 
3969 out:
3970 	free(fbuf, M_TEMP);
3971 	return (error);
3972 }
3973 
3974 /*
3975  * This is similar to vn_path_to_global_path but allows for regular
3976  * files which may not be present in the cache.
3977  *
3978  * Requires a locked, referenced vnode.
3979  * Vnode is re-locked on success or ENODEV, otherwise unlocked.
3980  */
3981 int
vn_path_to_global_path_hardlink(struct thread * td,struct vnode * vp,struct vnode * dvp,char * path,u_int pathlen,const char * leaf_name,size_t leaf_length)3982 vn_path_to_global_path_hardlink(struct thread *td, struct vnode *vp,
3983     struct vnode *dvp, char *path, u_int pathlen, const char *leaf_name,
3984     size_t leaf_length)
3985 {
3986 	struct nameidata nd;
3987 	struct vnode *vp1;
3988 	char *rpath, *fbuf;
3989 	size_t len;
3990 	int error;
3991 
3992 	ASSERT_VOP_ELOCKED(vp, __func__);
3993 
3994 	/*
3995 	 * Construct global filesystem path from dvp, vp and leaf
3996 	 * name.
3997 	 */
3998 	VOP_UNLOCK(vp);
3999 	len = pathlen;
4000 	error = vn_fullpath_hardlink(vp, dvp, leaf_name, leaf_length,
4001 	    &rpath, &fbuf, &len);
4002 
4003 	if (error != 0) {
4004 		vrele(vp);
4005 		return (error);
4006 	}
4007 
4008 	if (strlen(rpath) >= pathlen) {
4009 		vrele(vp);
4010 		error = ENAMETOOLONG;
4011 		goto out;
4012 	}
4013 
4014 	/*
4015 	 * Re-lookup the vnode by path to detect a possible rename.
4016 	 * As a side effect, the vnode is relocked.
4017 	 * If vnode was renamed, return ENOENT.
4018 	 */
4019 	NDINIT(&nd, LOOKUP, FOLLOW | LOCKLEAF | AUDITVNODE1, UIO_SYSSPACE, path);
4020 	error = namei(&nd);
4021 	if (error != 0) {
4022 		vrele(vp);
4023 		goto out;
4024 	}
4025 	NDFREE_PNBUF(&nd);
4026 	vp1 = nd.ni_vp;
4027 	vrele(vp);
4028 	if (vp1 == vp)
4029 		strcpy(path, rpath);
4030 	else {
4031 		vput(vp1);
4032 		error = ENOENT;
4033 	}
4034 
4035 out:
4036 	free(fbuf, M_TEMP);
4037 	return (error);
4038 }
4039 
4040 void
cache_vop_inotify(struct vnode * vp,int event,uint32_t cookie)4041 cache_vop_inotify(struct vnode *vp, int event, uint32_t cookie)
4042 {
4043 	struct mtx *vlp;
4044 	struct namecache *ncp;
4045 	int isdir;
4046 	bool logged, self;
4047 
4048 	isdir = vp->v_type == VDIR ? IN_ISDIR : 0;
4049 	self = (vn_irflag_read(vp) & VIRF_INOTIFY) != 0 &&
4050 	    (vp->v_type != VDIR || (event & ~_IN_DIR_EVENTS) != 0);
4051 
4052 	if (self) {
4053 		int selfevent;
4054 
4055 		if (event == _IN_ATTRIB_LINKCOUNT)
4056 			selfevent = IN_ATTRIB;
4057 		else
4058 			selfevent = event;
4059 		inotify_log(vp, NULL, 0, selfevent | isdir, cookie);
4060 	}
4061 	if ((event & IN_ALL_EVENTS) == 0)
4062 		return;
4063 
4064 	logged = false;
4065 	vlp = VP2VNODELOCK(vp);
4066 	mtx_lock(vlp);
4067 	TAILQ_FOREACH(ncp, &vp->v_cache_dst, nc_dst) {
4068 		if ((ncp->nc_flag & NCF_ISDOTDOT) != 0)
4069 			continue;
4070 		if ((vn_irflag_read(ncp->nc_dvp) & VIRF_INOTIFY) != 0) {
4071 			/*
4072 			 * XXX-MJ if the vnode has two links in the same
4073 			 * dir, we'll log the same event twice.
4074 			 */
4075 			inotify_log(ncp->nc_dvp, ncp->nc_name, ncp->nc_nlen,
4076 			    event | isdir, cookie);
4077 			logged = true;
4078 		}
4079 	}
4080 	if (!logged && (vn_irflag_read(vp) & VIRF_INOTIFY_PARENT) != 0) {
4081 		/*
4082 		 * We didn't find a watched directory that contains this vnode,
4083 		 * so stop calling VOP_INOTIFY for operations on the vnode.
4084 		 */
4085 		vn_irflag_unset(vp, VIRF_INOTIFY_PARENT);
4086 	}
4087 	mtx_unlock(vlp);
4088 }
4089 
4090 #ifdef DDB
4091 static void
db_print_vpath(struct vnode * vp)4092 db_print_vpath(struct vnode *vp)
4093 {
4094 
4095 	while (vp != NULL) {
4096 		db_printf("%p: ", vp);
4097 		if (vp == rootvnode) {
4098 			db_printf("/");
4099 			vp = NULL;
4100 		} else {
4101 			if (vp->v_vflag & VV_ROOT) {
4102 				db_printf("<mount point>");
4103 				vp = vp->v_mount->mnt_vnodecovered;
4104 			} else {
4105 				struct namecache *ncp;
4106 				char *ncn;
4107 				int i;
4108 
4109 				ncp = TAILQ_FIRST(&vp->v_cache_dst);
4110 				if (ncp != NULL) {
4111 					ncn = ncp->nc_name;
4112 					for (i = 0; i < ncp->nc_nlen; i++)
4113 						db_printf("%c", *ncn++);
4114 					vp = ncp->nc_dvp;
4115 				} else {
4116 					vp = NULL;
4117 				}
4118 			}
4119 		}
4120 		db_printf("\n");
4121 	}
4122 
4123 	return;
4124 }
4125 
DB_SHOW_COMMAND(vpath,db_show_vpath)4126 DB_SHOW_COMMAND(vpath, db_show_vpath)
4127 {
4128 	struct vnode *vp;
4129 
4130 	if (!have_addr) {
4131 		db_printf("usage: show vpath <struct vnode *>\n");
4132 		return;
4133 	}
4134 
4135 	vp = (struct vnode *)addr;
4136 	db_print_vpath(vp);
4137 }
4138 
4139 #endif
4140 
4141 static int cache_fast_lookup = 1;
4142 
4143 #define CACHE_FPL_FAILED	-2020
4144 
4145 static int
cache_vop_bad_vexec(struct vop_fplookup_vexec_args * v)4146 cache_vop_bad_vexec(struct vop_fplookup_vexec_args *v)
4147 {
4148 	vn_printf(v->a_vp, "no proper vop_fplookup_vexec\n");
4149 	panic("no proper vop_fplookup_vexec");
4150 }
4151 
4152 static int
cache_vop_bad_symlink(struct vop_fplookup_symlink_args * v)4153 cache_vop_bad_symlink(struct vop_fplookup_symlink_args *v)
4154 {
4155 	vn_printf(v->a_vp, "no proper vop_fplookup_symlink\n");
4156 	panic("no proper vop_fplookup_symlink");
4157 }
4158 
4159 void
cache_vop_vector_register(struct vop_vector * v)4160 cache_vop_vector_register(struct vop_vector *v)
4161 {
4162 	size_t ops;
4163 
4164 	ops = 0;
4165 	if (v->vop_fplookup_vexec != NULL) {
4166 		ops++;
4167 	}
4168 	if (v->vop_fplookup_symlink != NULL) {
4169 		ops++;
4170 	}
4171 
4172 	if (ops == 2) {
4173 		return;
4174 	}
4175 
4176 	if (ops == 0) {
4177 		v->vop_fplookup_vexec = cache_vop_bad_vexec;
4178 		v->vop_fplookup_symlink = cache_vop_bad_symlink;
4179 		return;
4180 	}
4181 
4182 	printf("%s: invalid vop vector %p -- either all or none fplookup vops "
4183 	    "need to be provided",  __func__, v);
4184 	if (v->vop_fplookup_vexec == NULL) {
4185 		printf("%s: missing vop_fplookup_vexec\n", __func__);
4186 	}
4187 	if (v->vop_fplookup_symlink == NULL) {
4188 		printf("%s: missing vop_fplookup_symlink\n", __func__);
4189 	}
4190 	panic("bad vop vector %p", v);
4191 }
4192 
4193 #ifdef INVARIANTS
4194 void
cache_validate_vop_vector(struct mount * mp,struct vop_vector * vops)4195 cache_validate_vop_vector(struct mount *mp, struct vop_vector *vops)
4196 {
4197 	if (mp == NULL)
4198 		return;
4199 
4200 	if ((mp->mnt_kern_flag & MNTK_FPLOOKUP) == 0)
4201 		return;
4202 
4203 	if (vops->vop_fplookup_vexec == NULL ||
4204 	    vops->vop_fplookup_vexec == cache_vop_bad_vexec)
4205 		panic("bad vop_fplookup_vexec on vector %p for filesystem %s",
4206 		    vops, mp->mnt_vfc->vfc_name);
4207 
4208 	if (vops->vop_fplookup_symlink == NULL ||
4209 	    vops->vop_fplookup_symlink == cache_vop_bad_symlink)
4210 		panic("bad vop_fplookup_symlink on vector %p for filesystem %s",
4211 		    vops, mp->mnt_vfc->vfc_name);
4212 }
4213 #endif
4214 
4215 void
cache_fast_lookup_enabled_recalc(void)4216 cache_fast_lookup_enabled_recalc(void)
4217 {
4218 	int lookup_flag;
4219 	int mac_on;
4220 
4221 #ifdef MAC
4222 	mac_on = mac_vnode_check_lookup_enabled();
4223 	mac_on |= mac_vnode_check_readlink_enabled();
4224 #else
4225 	mac_on = 0;
4226 #endif
4227 
4228 	lookup_flag = atomic_load_int(&cache_fast_lookup);
4229 	if (lookup_flag && !mac_on) {
4230 		atomic_store_char(&cache_fast_lookup_enabled, true);
4231 	} else {
4232 		atomic_store_char(&cache_fast_lookup_enabled, false);
4233 	}
4234 }
4235 
4236 static int
syscal_vfs_cache_fast_lookup(SYSCTL_HANDLER_ARGS)4237 syscal_vfs_cache_fast_lookup(SYSCTL_HANDLER_ARGS)
4238 {
4239 	int error, old;
4240 
4241 	old = atomic_load_int(&cache_fast_lookup);
4242 	error = sysctl_handle_int(oidp, arg1, arg2, req);
4243 	if (error == 0 && req->newptr && old != atomic_load_int(&cache_fast_lookup))
4244 		cache_fast_lookup_enabled_recalc();
4245 	return (error);
4246 }
4247 SYSCTL_PROC(_vfs_cache_param, OID_AUTO, fast_lookup, CTLTYPE_INT|CTLFLAG_RW|CTLFLAG_MPSAFE,
4248     &cache_fast_lookup, 0, syscal_vfs_cache_fast_lookup, "IU", "");
4249 
4250 /*
4251  * Components of nameidata (or objects it can point to) which may
4252  * need restoring in case fast path lookup fails.
4253  */
4254 struct nameidata_outer {
4255 	size_t ni_pathlen;
4256 	uint64_t cn_flags;
4257 };
4258 
4259 struct nameidata_saved {
4260 #ifdef INVARIANTS
4261 	char *cn_nameptr;
4262 	size_t ni_pathlen;
4263 #endif
4264 };
4265 
4266 #ifdef INVARIANTS
4267 struct cache_fpl_debug {
4268 	size_t ni_pathlen;
4269 };
4270 #endif
4271 
4272 struct cache_fpl {
4273 	struct nameidata *ndp;
4274 	struct componentname *cnp;
4275 	char *nulchar;
4276 	struct vnode *dvp;
4277 	struct vnode *tvp;
4278 	seqc_t dvp_seqc;
4279 	seqc_t tvp_seqc;
4280 	uint32_t hash;
4281 	struct nameidata_saved snd;
4282 	struct nameidata_outer snd_outer;
4283 	int line;
4284 	enum cache_fpl_status status:8;
4285 	bool in_smr;
4286 	bool fsearch;
4287 	struct pwd **pwd;
4288 #ifdef INVARIANTS
4289 	struct cache_fpl_debug debug;
4290 #endif
4291 };
4292 
4293 static bool cache_fplookup_mp_supported(struct mount *mp);
4294 static bool cache_fplookup_is_mp(struct cache_fpl *fpl);
4295 static int cache_fplookup_cross_mount(struct cache_fpl *fpl);
4296 static int cache_fplookup_partial_setup(struct cache_fpl *fpl);
4297 static int cache_fplookup_skip_slashes(struct cache_fpl *fpl);
4298 static int cache_fplookup_trailingslash(struct cache_fpl *fpl);
4299 static void cache_fpl_pathlen_dec(struct cache_fpl *fpl);
4300 static void cache_fpl_pathlen_inc(struct cache_fpl *fpl);
4301 static void cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n);
4302 static void cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n);
4303 
4304 static void
cache_fpl_cleanup_cnp(struct componentname * cnp)4305 cache_fpl_cleanup_cnp(struct componentname *cnp)
4306 {
4307 
4308 	uma_zfree(namei_zone, cnp->cn_pnbuf);
4309 	cnp->cn_pnbuf = NULL;
4310 	cnp->cn_nameptr = NULL;
4311 }
4312 
4313 static struct vnode *
cache_fpl_handle_root(struct cache_fpl * fpl)4314 cache_fpl_handle_root(struct cache_fpl *fpl)
4315 {
4316 	struct nameidata *ndp;
4317 	struct componentname *cnp;
4318 
4319 	ndp = fpl->ndp;
4320 	cnp = fpl->cnp;
4321 
4322 	MPASS(*(cnp->cn_nameptr) == '/');
4323 	cnp->cn_nameptr++;
4324 	cache_fpl_pathlen_dec(fpl);
4325 
4326 	if (__predict_false(*(cnp->cn_nameptr) == '/')) {
4327 		do {
4328 			cnp->cn_nameptr++;
4329 			cache_fpl_pathlen_dec(fpl);
4330 		} while (*(cnp->cn_nameptr) == '/');
4331 	}
4332 
4333 	return (ndp->ni_rootdir);
4334 }
4335 
4336 static void
cache_fpl_checkpoint_outer(struct cache_fpl * fpl)4337 cache_fpl_checkpoint_outer(struct cache_fpl *fpl)
4338 {
4339 
4340 	fpl->snd_outer.ni_pathlen = fpl->ndp->ni_pathlen;
4341 	fpl->snd_outer.cn_flags = fpl->ndp->ni_cnd.cn_flags;
4342 }
4343 
4344 static void
cache_fpl_checkpoint(struct cache_fpl * fpl)4345 cache_fpl_checkpoint(struct cache_fpl *fpl)
4346 {
4347 
4348 #ifdef INVARIANTS
4349 	fpl->snd.cn_nameptr = fpl->ndp->ni_cnd.cn_nameptr;
4350 	fpl->snd.ni_pathlen = fpl->debug.ni_pathlen;
4351 #endif
4352 }
4353 
4354 static void
cache_fpl_restore_partial(struct cache_fpl * fpl)4355 cache_fpl_restore_partial(struct cache_fpl *fpl)
4356 {
4357 
4358 	fpl->ndp->ni_cnd.cn_flags = fpl->snd_outer.cn_flags;
4359 #ifdef INVARIANTS
4360 	fpl->debug.ni_pathlen = fpl->snd.ni_pathlen;
4361 #endif
4362 }
4363 
4364 static void
cache_fpl_restore_abort(struct cache_fpl * fpl)4365 cache_fpl_restore_abort(struct cache_fpl *fpl)
4366 {
4367 
4368 	cache_fpl_restore_partial(fpl);
4369 	/*
4370 	 * It is 0 on entry by API contract.
4371 	 */
4372 	fpl->ndp->ni_resflags = 0;
4373 	fpl->ndp->ni_cnd.cn_nameptr = fpl->ndp->ni_cnd.cn_pnbuf;
4374 	fpl->ndp->ni_pathlen = fpl->snd_outer.ni_pathlen;
4375 }
4376 
4377 #ifdef INVARIANTS
4378 #define cache_fpl_smr_assert_entered(fpl) ({			\
4379 	struct cache_fpl *_fpl = (fpl);				\
4380 	MPASS(_fpl->in_smr == true);				\
4381 	VFS_SMR_ASSERT_ENTERED();				\
4382 })
4383 #define cache_fpl_smr_assert_not_entered(fpl) ({		\
4384 	struct cache_fpl *_fpl = (fpl);				\
4385 	MPASS(_fpl->in_smr == false);				\
4386 	VFS_SMR_ASSERT_NOT_ENTERED();				\
4387 })
4388 static void
cache_fpl_assert_status(struct cache_fpl * fpl)4389 cache_fpl_assert_status(struct cache_fpl *fpl)
4390 {
4391 
4392 	switch (fpl->status) {
4393 	case CACHE_FPL_STATUS_UNSET:
4394 		__assert_unreachable();
4395 		break;
4396 	case CACHE_FPL_STATUS_DESTROYED:
4397 	case CACHE_FPL_STATUS_ABORTED:
4398 	case CACHE_FPL_STATUS_PARTIAL:
4399 	case CACHE_FPL_STATUS_HANDLED:
4400 		break;
4401 	}
4402 }
4403 #else
4404 #define cache_fpl_smr_assert_entered(fpl) do { } while (0)
4405 #define cache_fpl_smr_assert_not_entered(fpl) do { } while (0)
4406 #define cache_fpl_assert_status(fpl) do { } while (0)
4407 #endif
4408 
4409 #define cache_fpl_smr_enter_initial(fpl) ({			\
4410 	struct cache_fpl *_fpl = (fpl);				\
4411 	vfs_smr_enter();					\
4412 	_fpl->in_smr = true;					\
4413 })
4414 
4415 #define cache_fpl_smr_enter(fpl) ({				\
4416 	struct cache_fpl *_fpl = (fpl);				\
4417 	MPASS(_fpl->in_smr == false);				\
4418 	vfs_smr_enter();					\
4419 	_fpl->in_smr = true;					\
4420 })
4421 
4422 #define cache_fpl_smr_exit(fpl) ({				\
4423 	struct cache_fpl *_fpl = (fpl);				\
4424 	MPASS(_fpl->in_smr == true);				\
4425 	vfs_smr_exit();						\
4426 	_fpl->in_smr = false;					\
4427 })
4428 
4429 static int
cache_fpl_aborted_early_impl(struct cache_fpl * fpl,int line)4430 cache_fpl_aborted_early_impl(struct cache_fpl *fpl, int line)
4431 {
4432 
4433 	if (fpl->status != CACHE_FPL_STATUS_UNSET) {
4434 		KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL,
4435 		    ("%s: converting to abort from %d at %d, set at %d\n",
4436 		    __func__, fpl->status, line, fpl->line));
4437 	}
4438 	cache_fpl_smr_assert_not_entered(fpl);
4439 	fpl->status = CACHE_FPL_STATUS_ABORTED;
4440 	fpl->line = line;
4441 	return (CACHE_FPL_FAILED);
4442 }
4443 
4444 #define cache_fpl_aborted_early(x)	cache_fpl_aborted_early_impl((x), __LINE__)
4445 
4446 static int __noinline
cache_fpl_aborted_impl(struct cache_fpl * fpl,int line)4447 cache_fpl_aborted_impl(struct cache_fpl *fpl, int line)
4448 {
4449 	struct nameidata *ndp;
4450 	struct componentname *cnp;
4451 
4452 	ndp = fpl->ndp;
4453 	cnp = fpl->cnp;
4454 
4455 	if (fpl->status != CACHE_FPL_STATUS_UNSET) {
4456 		KASSERT(fpl->status == CACHE_FPL_STATUS_PARTIAL,
4457 		    ("%s: converting to abort from %d at %d, set at %d\n",
4458 		    __func__, fpl->status, line, fpl->line));
4459 	}
4460 	fpl->status = CACHE_FPL_STATUS_ABORTED;
4461 	fpl->line = line;
4462 	if (fpl->in_smr)
4463 		cache_fpl_smr_exit(fpl);
4464 	cache_fpl_restore_abort(fpl);
4465 	/*
4466 	 * Resolving symlinks overwrites data passed by the caller.
4467 	 * Let namei know.
4468 	 */
4469 	if (ndp->ni_loopcnt > 0) {
4470 		fpl->status = CACHE_FPL_STATUS_DESTROYED;
4471 		cache_fpl_cleanup_cnp(cnp);
4472 	}
4473 	return (CACHE_FPL_FAILED);
4474 }
4475 
4476 #define cache_fpl_aborted(x)	cache_fpl_aborted_impl((x), __LINE__)
4477 
4478 static int __noinline
cache_fpl_partial_impl(struct cache_fpl * fpl,int line)4479 cache_fpl_partial_impl(struct cache_fpl *fpl, int line)
4480 {
4481 
4482 	KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
4483 	    ("%s: setting to partial at %d, but already set to %d at %d\n",
4484 	    __func__, line, fpl->status, fpl->line));
4485 	cache_fpl_smr_assert_entered(fpl);
4486 	fpl->status = CACHE_FPL_STATUS_PARTIAL;
4487 	fpl->line = line;
4488 	return (cache_fplookup_partial_setup(fpl));
4489 }
4490 
4491 #define cache_fpl_partial(x)	cache_fpl_partial_impl((x), __LINE__)
4492 
4493 static int
cache_fpl_handled_impl(struct cache_fpl * fpl,int line)4494 cache_fpl_handled_impl(struct cache_fpl *fpl, int line)
4495 {
4496 
4497 	KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
4498 	    ("%s: setting to handled at %d, but already set to %d at %d\n",
4499 	    __func__, line, fpl->status, fpl->line));
4500 	cache_fpl_smr_assert_not_entered(fpl);
4501 	fpl->status = CACHE_FPL_STATUS_HANDLED;
4502 	fpl->line = line;
4503 	return (0);
4504 }
4505 
4506 #define cache_fpl_handled(x)	cache_fpl_handled_impl((x), __LINE__)
4507 
4508 static int
cache_fpl_handled_error_impl(struct cache_fpl * fpl,int error,int line)4509 cache_fpl_handled_error_impl(struct cache_fpl *fpl, int error, int line)
4510 {
4511 
4512 	KASSERT(fpl->status == CACHE_FPL_STATUS_UNSET,
4513 	    ("%s: setting to handled at %d, but already set to %d at %d\n",
4514 	    __func__, line, fpl->status, fpl->line));
4515 	MPASS(error != 0);
4516 	MPASS(error != CACHE_FPL_FAILED);
4517 	cache_fpl_smr_assert_not_entered(fpl);
4518 	fpl->status = CACHE_FPL_STATUS_HANDLED;
4519 	fpl->line = line;
4520 	fpl->dvp = NULL;
4521 	fpl->tvp = NULL;
4522 	return (error);
4523 }
4524 
4525 #define cache_fpl_handled_error(x, e)	cache_fpl_handled_error_impl((x), (e), __LINE__)
4526 
4527 static bool
cache_fpl_terminated(struct cache_fpl * fpl)4528 cache_fpl_terminated(struct cache_fpl *fpl)
4529 {
4530 
4531 	return (fpl->status != CACHE_FPL_STATUS_UNSET);
4532 }
4533 
4534 #define CACHE_FPL_SUPPORTED_CN_FLAGS \
4535 	(NC_NOMAKEENTRY | NC_KEEPPOSENTRY | LOCKLEAF | LOCKPARENT | WANTPARENT | \
4536 	 FAILIFEXISTS | FOLLOW | EMPTYPATH | LOCKSHARED | ISRESTARTED | WILLBEDIR | \
4537 	 ISOPEN | NOMACCHECK | AUDITVNODE1 | AUDITVNODE2 | NOCAPCHECK | OPENREAD | \
4538 	 OPENWRITE | WANTIOCTLCAPS | NAMEILOOKUP)
4539 
4540 #define CACHE_FPL_INTERNAL_CN_FLAGS \
4541 	(ISDOTDOT | MAKEENTRY | ISLASTCN)
4542 
4543 _Static_assert((CACHE_FPL_SUPPORTED_CN_FLAGS & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
4544     "supported and internal flags overlap");
4545 
4546 static bool
cache_fpl_islastcn(struct nameidata * ndp)4547 cache_fpl_islastcn(struct nameidata *ndp)
4548 {
4549 
4550 	return (*ndp->ni_next == 0);
4551 }
4552 
4553 static bool
cache_fpl_istrailingslash(struct cache_fpl * fpl)4554 cache_fpl_istrailingslash(struct cache_fpl *fpl)
4555 {
4556 
4557 	MPASS(fpl->nulchar > fpl->cnp->cn_pnbuf);
4558 	return (*(fpl->nulchar - 1) == '/');
4559 }
4560 
4561 static bool
cache_fpl_isdotdot(struct componentname * cnp)4562 cache_fpl_isdotdot(struct componentname *cnp)
4563 {
4564 
4565 	if (cnp->cn_namelen == 2 &&
4566 	    cnp->cn_nameptr[1] == '.' && cnp->cn_nameptr[0] == '.')
4567 		return (true);
4568 	return (false);
4569 }
4570 
4571 static bool
cache_can_fplookup(struct cache_fpl * fpl)4572 cache_can_fplookup(struct cache_fpl *fpl)
4573 {
4574 	struct nameidata *ndp;
4575 	struct componentname *cnp;
4576 	struct thread *td;
4577 
4578 	ndp = fpl->ndp;
4579 	cnp = fpl->cnp;
4580 	td = curthread;
4581 
4582 	if (!atomic_load_char(&cache_fast_lookup_enabled)) {
4583 		cache_fpl_aborted_early(fpl);
4584 		return (false);
4585 	}
4586 	if ((cnp->cn_flags & ~CACHE_FPL_SUPPORTED_CN_FLAGS) != 0) {
4587 		cache_fpl_aborted_early(fpl);
4588 		return (false);
4589 	}
4590 	if (IN_CAPABILITY_MODE(td) || CAP_TRACING(td)) {
4591 		cache_fpl_aborted_early(fpl);
4592 		return (false);
4593 	}
4594 	if (AUDITING_TD(td)) {
4595 		cache_fpl_aborted_early(fpl);
4596 		return (false);
4597 	}
4598 	if (ndp->ni_startdir != NULL) {
4599 		cache_fpl_aborted_early(fpl);
4600 		return (false);
4601 	}
4602 	return (true);
4603 }
4604 
4605 static int __noinline
cache_fplookup_dirfd(struct cache_fpl * fpl,struct vnode ** vpp)4606 cache_fplookup_dirfd(struct cache_fpl *fpl, struct vnode **vpp)
4607 {
4608 	struct nameidata *ndp;
4609 	struct componentname *cnp;
4610 	int error, flags;
4611 
4612 	ndp = fpl->ndp;
4613 	cnp = fpl->cnp;
4614 
4615 	error = fgetvp_lookup_smr(ndp, vpp, &flags);
4616 	if (__predict_false(error != 0)) {
4617 		return (cache_fpl_aborted(fpl));
4618 	}
4619 	if (__predict_false((flags & O_RESOLVE_BENEATH) != 0)) {
4620 		_Static_assert((CACHE_FPL_SUPPORTED_CN_FLAGS & RBENEATH) == 0,
4621 		    "RBENEATH supported by fplookup");
4622 		cache_fpl_smr_exit(fpl);
4623 		cache_fpl_aborted(fpl);
4624 		return (EOPNOTSUPP);
4625 	}
4626 	fpl->fsearch = (flags & FSEARCH) != 0;
4627 	if ((*vpp)->v_type != VDIR) {
4628 		if (!((cnp->cn_flags & EMPTYPATH) != 0 && cnp->cn_pnbuf[0] == '\0')) {
4629 			cache_fpl_smr_exit(fpl);
4630 			return (cache_fpl_handled_error(fpl, ENOTDIR));
4631 		}
4632 	}
4633 	return (0);
4634 }
4635 
4636 static int __noinline
cache_fplookup_negative_promote(struct cache_fpl * fpl,struct namecache * oncp,uint32_t hash)4637 cache_fplookup_negative_promote(struct cache_fpl *fpl, struct namecache *oncp,
4638     uint32_t hash)
4639 {
4640 	struct componentname *cnp;
4641 	struct vnode *dvp;
4642 
4643 	cnp = fpl->cnp;
4644 	dvp = fpl->dvp;
4645 
4646 	cache_fpl_smr_exit(fpl);
4647 	if (cache_neg_promote_cond(dvp, cnp, oncp, hash))
4648 		return (cache_fpl_handled_error(fpl, ENOENT));
4649 	else
4650 		return (cache_fpl_aborted(fpl));
4651 }
4652 
4653 /*
4654  * Prepare fallback to the locked lookup while trying to retain the progress.
4655  */
4656 static int __noinline
cache_fplookup_partial_setup(struct cache_fpl * fpl)4657 cache_fplookup_partial_setup(struct cache_fpl *fpl)
4658 {
4659 	struct nameidata *ndp;
4660 	struct componentname *cnp;
4661 	enum vgetstate dvs;
4662 	struct vnode *dvp;
4663 	struct pwd *pwd;
4664 	seqc_t dvp_seqc;
4665 
4666 	ndp = fpl->ndp;
4667 	cnp = fpl->cnp;
4668 	pwd = *(fpl->pwd);
4669 	dvp = fpl->dvp;
4670 	dvp_seqc = fpl->dvp_seqc;
4671 
4672 	if (!pwd_hold_smr(pwd)) {
4673 		return (cache_fpl_aborted(fpl));
4674 	}
4675 
4676 	/*
4677 	 * Note that seqc is checked before the vnode is locked, so by
4678 	 * the time regular lookup gets to it it may have moved.
4679 	 *
4680 	 * Ultimately this does not affect correctness, any lookup errors
4681 	 * are userspace racing with itself. It is guaranteed that any
4682 	 * path which ultimately gets found could also have been found
4683 	 * by regular lookup going all the way in absence of concurrent
4684 	 * modifications.
4685 	 */
4686 	dvs = vget_prep_smr(dvp);
4687 	cache_fpl_smr_exit(fpl);
4688 	if (__predict_false(dvs == VGET_NONE)) {
4689 		pwd_drop(pwd);
4690 		return (cache_fpl_aborted(fpl));
4691 	}
4692 
4693 	vget_finish_ref(dvp, dvs);
4694 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4695 		vrele(dvp);
4696 		pwd_drop(pwd);
4697 		return (cache_fpl_aborted(fpl));
4698 	}
4699 
4700 	cache_fpl_restore_partial(fpl);
4701 #ifdef INVARIANTS
4702 	if (cnp->cn_nameptr != fpl->snd.cn_nameptr) {
4703 		panic("%s: cn_nameptr mismatch (%p != %p) full [%s]\n", __func__,
4704 		    cnp->cn_nameptr, fpl->snd.cn_nameptr, cnp->cn_pnbuf);
4705 	}
4706 #endif
4707 
4708 	ndp->ni_startdir = dvp;
4709 	cnp->cn_flags |= MAKEENTRY;
4710 	if (cache_fpl_islastcn(ndp))
4711 		cnp->cn_flags |= ISLASTCN;
4712 	if (cache_fpl_isdotdot(cnp))
4713 		cnp->cn_flags |= ISDOTDOT;
4714 
4715 	/*
4716 	 * Skip potential extra slashes parsing did not take care of.
4717 	 * cache_fplookup_skip_slashes explains the mechanism.
4718 	 */
4719 	if (__predict_false(*(cnp->cn_nameptr) == '/')) {
4720 		do {
4721 			cnp->cn_nameptr++;
4722 			cache_fpl_pathlen_dec(fpl);
4723 		} while (*(cnp->cn_nameptr) == '/');
4724 	}
4725 
4726 	ndp->ni_pathlen = fpl->nulchar - cnp->cn_nameptr + 1;
4727 #ifdef INVARIANTS
4728 	if (ndp->ni_pathlen != fpl->debug.ni_pathlen) {
4729 		panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n",
4730 		    __func__, ndp->ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar,
4731 		    cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf);
4732 	}
4733 #endif
4734 	return (0);
4735 }
4736 
4737 static int
cache_fplookup_final_child(struct cache_fpl * fpl,enum vgetstate tvs)4738 cache_fplookup_final_child(struct cache_fpl *fpl, enum vgetstate tvs)
4739 {
4740 	struct componentname *cnp;
4741 	struct vnode *tvp;
4742 	seqc_t tvp_seqc;
4743 	int error, lkflags;
4744 
4745 	cnp = fpl->cnp;
4746 	tvp = fpl->tvp;
4747 	tvp_seqc = fpl->tvp_seqc;
4748 
4749 	if ((cnp->cn_flags & LOCKLEAF) != 0) {
4750 		lkflags = LK_SHARED;
4751 		if ((cnp->cn_flags & LOCKSHARED) == 0)
4752 			lkflags = LK_EXCLUSIVE;
4753 		error = vget_finish(tvp, lkflags, tvs);
4754 		if (__predict_false(error != 0)) {
4755 			return (cache_fpl_aborted(fpl));
4756 		}
4757 	} else {
4758 		vget_finish_ref(tvp, tvs);
4759 	}
4760 
4761 	if (!vn_seqc_consistent(tvp, tvp_seqc)) {
4762 		if ((cnp->cn_flags & LOCKLEAF) != 0)
4763 			vput(tvp);
4764 		else
4765 			vrele(tvp);
4766 		return (cache_fpl_aborted(fpl));
4767 	}
4768 
4769 	return (cache_fpl_handled(fpl));
4770 }
4771 
4772 /*
4773  * They want to possibly modify the state of the namecache.
4774  */
4775 static int __noinline
cache_fplookup_final_modifying(struct cache_fpl * fpl)4776 cache_fplookup_final_modifying(struct cache_fpl *fpl)
4777 {
4778 	struct nameidata *ndp __diagused;
4779 	struct componentname *cnp;
4780 	enum vgetstate dvs;
4781 	struct vnode *dvp, *tvp;
4782 	struct mount *mp;
4783 	seqc_t dvp_seqc;
4784 	int error;
4785 	bool docache;
4786 
4787 	ndp = fpl->ndp;
4788 	cnp = fpl->cnp;
4789 	dvp = fpl->dvp;
4790 	dvp_seqc = fpl->dvp_seqc;
4791 
4792 	MPASS(*(cnp->cn_nameptr) != '/');
4793 	MPASS(cache_fpl_islastcn(ndp));
4794 	if ((cnp->cn_flags & LOCKPARENT) == 0)
4795 		MPASS((cnp->cn_flags & WANTPARENT) != 0);
4796 	MPASS((cnp->cn_flags & TRAILINGSLASH) == 0);
4797 	MPASS(cnp->cn_nameiop == CREATE || cnp->cn_nameiop == DELETE ||
4798 	    cnp->cn_nameiop == RENAME);
4799 	MPASS((cnp->cn_flags & MAKEENTRY) == 0);
4800 	MPASS((cnp->cn_flags & ISDOTDOT) == 0);
4801 
4802 	docache = (cnp->cn_flags & NOCACHE) ^ NOCACHE;
4803 	if (cnp->cn_nameiop == DELETE || cnp->cn_nameiop == RENAME)
4804 		docache = false;
4805 
4806 	/*
4807 	 * Regular lookup nulifies the slash, which we don't do here.
4808 	 * Don't take chances with filesystem routines seeing it for
4809 	 * the last entry.
4810 	 */
4811 	if (cache_fpl_istrailingslash(fpl)) {
4812 		return (cache_fpl_partial(fpl));
4813 	}
4814 
4815 	mp = atomic_load_ptr(&dvp->v_mount);
4816 	if (__predict_false(mp == NULL)) {
4817 		return (cache_fpl_aborted(fpl));
4818 	}
4819 
4820 	if (__predict_false(mp->mnt_flag & MNT_RDONLY)) {
4821 		cache_fpl_smr_exit(fpl);
4822 		/*
4823 		 * Original code keeps not checking for CREATE which
4824 		 * might be a bug. For now let the old lookup decide.
4825 		 */
4826 		if (cnp->cn_nameiop == CREATE) {
4827 			return (cache_fpl_aborted(fpl));
4828 		}
4829 		return (cache_fpl_handled_error(fpl, EROFS));
4830 	}
4831 
4832 	if (fpl->tvp != NULL && (cnp->cn_flags & FAILIFEXISTS) != 0) {
4833 		cache_fpl_smr_exit(fpl);
4834 		return (cache_fpl_handled_error(fpl, EEXIST));
4835 	}
4836 
4837 	/*
4838 	 * Secure access to dvp; check cache_fplookup_partial_setup for
4839 	 * reasoning.
4840 	 *
4841 	 * XXX At least UFS requires its lookup routine to be called for
4842 	 * the last path component, which leads to some level of complication
4843 	 * and inefficiency:
4844 	 * - the target routine always locks the target vnode, but our caller
4845 	 *   may not need it locked
4846 	 * - some of the VOP machinery asserts that the parent is locked, which
4847 	 *   once more may be not required
4848 	 *
4849 	 * TODO: add a flag for filesystems which don't need this.
4850 	 */
4851 	dvs = vget_prep_smr(dvp);
4852 	cache_fpl_smr_exit(fpl);
4853 	if (__predict_false(dvs == VGET_NONE)) {
4854 		return (cache_fpl_aborted(fpl));
4855 	}
4856 
4857 	vget_finish_ref(dvp, dvs);
4858 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
4859 		vrele(dvp);
4860 		return (cache_fpl_aborted(fpl));
4861 	}
4862 
4863 	error = vn_lock(dvp, LK_EXCLUSIVE);
4864 	if (__predict_false(error != 0)) {
4865 		vrele(dvp);
4866 		return (cache_fpl_aborted(fpl));
4867 	}
4868 
4869 	tvp = NULL;
4870 	cnp->cn_flags |= ISLASTCN;
4871 	if (docache)
4872 		cnp->cn_flags |= MAKEENTRY;
4873 	if (cache_fpl_isdotdot(cnp))
4874 		cnp->cn_flags |= ISDOTDOT;
4875 	cnp->cn_lkflags = LK_EXCLUSIVE;
4876 	error = VOP_LOOKUP(dvp, &tvp, cnp);
4877 	switch (error) {
4878 	case EJUSTRETURN:
4879 	case 0:
4880 		break;
4881 	case ENOTDIR:
4882 	case ENOENT:
4883 		vput(dvp);
4884 		return (cache_fpl_handled_error(fpl, error));
4885 	default:
4886 		vput(dvp);
4887 		return (cache_fpl_aborted(fpl));
4888 	}
4889 
4890 	fpl->tvp = tvp;
4891 
4892 	if (tvp == NULL) {
4893 		MPASS(error == EJUSTRETURN);
4894 		if ((cnp->cn_flags & LOCKPARENT) == 0) {
4895 			VOP_UNLOCK(dvp);
4896 		}
4897 		return (cache_fpl_handled(fpl));
4898 	}
4899 
4900 	/*
4901 	 * There are very hairy corner cases concerning various flag combinations
4902 	 * and locking state. In particular here we only hold one lock instead of
4903 	 * two.
4904 	 *
4905 	 * Skip the complexity as it is of no significance for normal workloads.
4906 	 */
4907 	if (__predict_false(tvp == dvp)) {
4908 		vput(dvp);
4909 		vrele(tvp);
4910 		return (cache_fpl_aborted(fpl));
4911 	}
4912 
4913 	/*
4914 	 * If they want the symlink itself we are fine, but if they want to
4915 	 * follow it regular lookup has to be engaged.
4916 	 */
4917 	if (tvp->v_type == VLNK) {
4918 		if ((cnp->cn_flags & FOLLOW) != 0) {
4919 			vput(dvp);
4920 			vput(tvp);
4921 			return (cache_fpl_aborted(fpl));
4922 		}
4923 	}
4924 
4925 	/*
4926 	 * Since we expect this to be the terminal vnode it should almost never
4927 	 * be a mount point.
4928 	 */
4929 	if (__predict_false(cache_fplookup_is_mp(fpl))) {
4930 		vput(dvp);
4931 		vput(tvp);
4932 		return (cache_fpl_aborted(fpl));
4933 	}
4934 
4935 	if ((cnp->cn_flags & FAILIFEXISTS) != 0) {
4936 		vput(dvp);
4937 		vput(tvp);
4938 		return (cache_fpl_handled_error(fpl, EEXIST));
4939 	}
4940 
4941 	if ((cnp->cn_flags & LOCKLEAF) == 0) {
4942 		VOP_UNLOCK(tvp);
4943 	}
4944 
4945 	if ((cnp->cn_flags & LOCKPARENT) == 0) {
4946 		VOP_UNLOCK(dvp);
4947 	}
4948 
4949 	return (cache_fpl_handled(fpl));
4950 }
4951 
4952 static int __noinline
cache_fplookup_modifying(struct cache_fpl * fpl)4953 cache_fplookup_modifying(struct cache_fpl *fpl)
4954 {
4955 	struct nameidata *ndp;
4956 
4957 	ndp = fpl->ndp;
4958 
4959 	if (!cache_fpl_islastcn(ndp)) {
4960 		return (cache_fpl_partial(fpl));
4961 	}
4962 	return (cache_fplookup_final_modifying(fpl));
4963 }
4964 
4965 static int __noinline
cache_fplookup_final_withparent(struct cache_fpl * fpl)4966 cache_fplookup_final_withparent(struct cache_fpl *fpl)
4967 {
4968 	struct componentname *cnp;
4969 	enum vgetstate dvs, tvs;
4970 	struct vnode *dvp, *tvp;
4971 	seqc_t dvp_seqc;
4972 	int error;
4973 
4974 	cnp = fpl->cnp;
4975 	dvp = fpl->dvp;
4976 	dvp_seqc = fpl->dvp_seqc;
4977 	tvp = fpl->tvp;
4978 
4979 	MPASS((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0);
4980 
4981 	/*
4982 	 * This is less efficient than it can be for simplicity.
4983 	 */
4984 	dvs = vget_prep_smr(dvp);
4985 	if (__predict_false(dvs == VGET_NONE)) {
4986 		return (cache_fpl_aborted(fpl));
4987 	}
4988 	tvs = vget_prep_smr(tvp);
4989 	if (__predict_false(tvs == VGET_NONE)) {
4990 		cache_fpl_smr_exit(fpl);
4991 		vget_abort(dvp, dvs);
4992 		return (cache_fpl_aborted(fpl));
4993 	}
4994 
4995 	cache_fpl_smr_exit(fpl);
4996 
4997 	if ((cnp->cn_flags & LOCKPARENT) != 0) {
4998 		error = vget_finish(dvp, LK_EXCLUSIVE, dvs);
4999 		if (__predict_false(error != 0)) {
5000 			vget_abort(tvp, tvs);
5001 			return (cache_fpl_aborted(fpl));
5002 		}
5003 	} else {
5004 		vget_finish_ref(dvp, dvs);
5005 	}
5006 
5007 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
5008 		vget_abort(tvp, tvs);
5009 		if ((cnp->cn_flags & LOCKPARENT) != 0)
5010 			vput(dvp);
5011 		else
5012 			vrele(dvp);
5013 		return (cache_fpl_aborted(fpl));
5014 	}
5015 
5016 	error = cache_fplookup_final_child(fpl, tvs);
5017 	if (__predict_false(error != 0)) {
5018 		MPASS(fpl->status == CACHE_FPL_STATUS_ABORTED ||
5019 		    fpl->status == CACHE_FPL_STATUS_DESTROYED);
5020 		if ((cnp->cn_flags & LOCKPARENT) != 0)
5021 			vput(dvp);
5022 		else
5023 			vrele(dvp);
5024 		return (error);
5025 	}
5026 
5027 	MPASS(fpl->status == CACHE_FPL_STATUS_HANDLED);
5028 	return (0);
5029 }
5030 
5031 static int
cache_fplookup_final(struct cache_fpl * fpl)5032 cache_fplookup_final(struct cache_fpl *fpl)
5033 {
5034 	struct componentname *cnp;
5035 	enum vgetstate tvs;
5036 	struct vnode *dvp, *tvp;
5037 	seqc_t dvp_seqc;
5038 
5039 	cnp = fpl->cnp;
5040 	dvp = fpl->dvp;
5041 	dvp_seqc = fpl->dvp_seqc;
5042 	tvp = fpl->tvp;
5043 
5044 	MPASS(*(cnp->cn_nameptr) != '/');
5045 
5046 	if (cnp->cn_nameiop != LOOKUP) {
5047 		return (cache_fplookup_final_modifying(fpl));
5048 	}
5049 
5050 	if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0)
5051 		return (cache_fplookup_final_withparent(fpl));
5052 
5053 	tvs = vget_prep_smr(tvp);
5054 	if (__predict_false(tvs == VGET_NONE)) {
5055 		return (cache_fpl_partial(fpl));
5056 	}
5057 
5058 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
5059 		cache_fpl_smr_exit(fpl);
5060 		vget_abort(tvp, tvs);
5061 		return (cache_fpl_aborted(fpl));
5062 	}
5063 
5064 	cache_fpl_smr_exit(fpl);
5065 	return (cache_fplookup_final_child(fpl, tvs));
5066 }
5067 
5068 /*
5069  * Comment from locked lookup:
5070  * Check for degenerate name (e.g. / or "") which is a way of talking about a
5071  * directory, e.g. like "/." or ".".
5072  */
5073 static int __noinline
cache_fplookup_degenerate(struct cache_fpl * fpl)5074 cache_fplookup_degenerate(struct cache_fpl *fpl)
5075 {
5076 	struct componentname *cnp;
5077 	struct vnode *dvp;
5078 	enum vgetstate dvs;
5079 	int error, lkflags;
5080 #ifdef INVARIANTS
5081 	char *cp;
5082 #endif
5083 
5084 	fpl->tvp = fpl->dvp;
5085 	fpl->tvp_seqc = fpl->dvp_seqc;
5086 
5087 	cnp = fpl->cnp;
5088 	dvp = fpl->dvp;
5089 
5090 #ifdef INVARIANTS
5091 	for (cp = cnp->cn_pnbuf; *cp != '\0'; cp++) {
5092 		KASSERT(*cp == '/',
5093 		    ("%s: encountered non-slash; string [%s]\n", __func__,
5094 		    cnp->cn_pnbuf));
5095 	}
5096 #endif
5097 
5098 	if (__predict_false(cnp->cn_nameiop != LOOKUP)) {
5099 		cache_fpl_smr_exit(fpl);
5100 		return (cache_fpl_handled_error(fpl, EISDIR));
5101 	}
5102 
5103 	if ((cnp->cn_flags & (LOCKPARENT|WANTPARENT)) != 0) {
5104 		return (cache_fplookup_final_withparent(fpl));
5105 	}
5106 
5107 	dvs = vget_prep_smr(dvp);
5108 	cache_fpl_smr_exit(fpl);
5109 	if (__predict_false(dvs == VGET_NONE)) {
5110 		return (cache_fpl_aborted(fpl));
5111 	}
5112 
5113 	if ((cnp->cn_flags & LOCKLEAF) != 0) {
5114 		lkflags = LK_SHARED;
5115 		if ((cnp->cn_flags & LOCKSHARED) == 0)
5116 			lkflags = LK_EXCLUSIVE;
5117 		error = vget_finish(dvp, lkflags, dvs);
5118 		if (__predict_false(error != 0)) {
5119 			return (cache_fpl_aborted(fpl));
5120 		}
5121 	} else {
5122 		vget_finish_ref(dvp, dvs);
5123 	}
5124 	return (cache_fpl_handled(fpl));
5125 }
5126 
5127 static int __noinline
cache_fplookup_emptypath(struct cache_fpl * fpl)5128 cache_fplookup_emptypath(struct cache_fpl *fpl)
5129 {
5130 	struct nameidata *ndp;
5131 	struct componentname *cnp;
5132 	enum vgetstate tvs;
5133 	struct vnode *tvp;
5134 	int error, lkflags;
5135 
5136 	fpl->tvp = fpl->dvp;
5137 	fpl->tvp_seqc = fpl->dvp_seqc;
5138 
5139 	ndp = fpl->ndp;
5140 	cnp = fpl->cnp;
5141 	tvp = fpl->tvp;
5142 
5143 	MPASS(*cnp->cn_pnbuf == '\0');
5144 
5145 	if (__predict_false((cnp->cn_flags & EMPTYPATH) == 0)) {
5146 		cache_fpl_smr_exit(fpl);
5147 		return (cache_fpl_handled_error(fpl, ENOENT));
5148 	}
5149 
5150 	MPASS((cnp->cn_flags & (LOCKPARENT | WANTPARENT)) == 0);
5151 
5152 	tvs = vget_prep_smr(tvp);
5153 	cache_fpl_smr_exit(fpl);
5154 	if (__predict_false(tvs == VGET_NONE)) {
5155 		return (cache_fpl_aborted(fpl));
5156 	}
5157 
5158 	if ((cnp->cn_flags & LOCKLEAF) != 0) {
5159 		lkflags = LK_SHARED;
5160 		if ((cnp->cn_flags & LOCKSHARED) == 0)
5161 			lkflags = LK_EXCLUSIVE;
5162 		error = vget_finish(tvp, lkflags, tvs);
5163 		if (__predict_false(error != 0)) {
5164 			return (cache_fpl_aborted(fpl));
5165 		}
5166 	} else {
5167 		vget_finish_ref(tvp, tvs);
5168 	}
5169 
5170 	ndp->ni_resflags |= NIRES_EMPTYPATH;
5171 	return (cache_fpl_handled(fpl));
5172 }
5173 
5174 static int __noinline
cache_fplookup_noentry(struct cache_fpl * fpl)5175 cache_fplookup_noentry(struct cache_fpl *fpl)
5176 {
5177 	struct nameidata *ndp;
5178 	struct componentname *cnp;
5179 	enum vgetstate dvs;
5180 	struct vnode *dvp, *tvp;
5181 	seqc_t dvp_seqc;
5182 	int error;
5183 
5184 	ndp = fpl->ndp;
5185 	cnp = fpl->cnp;
5186 	dvp = fpl->dvp;
5187 	dvp_seqc = fpl->dvp_seqc;
5188 
5189 	MPASS((cnp->cn_flags & MAKEENTRY) == 0);
5190 	MPASS((cnp->cn_flags & ISDOTDOT) == 0);
5191 	if (cnp->cn_nameiop == LOOKUP)
5192 		MPASS((cnp->cn_flags & NOCACHE) == 0);
5193 	MPASS(!cache_fpl_isdotdot(cnp));
5194 
5195 	/*
5196 	 * Hack: delayed name len checking.
5197 	 */
5198 	if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
5199 		cache_fpl_smr_exit(fpl);
5200 		return (cache_fpl_handled_error(fpl, ENAMETOOLONG));
5201 	}
5202 
5203 	if (cnp->cn_nameptr[0] == '/') {
5204 		return (cache_fplookup_skip_slashes(fpl));
5205 	}
5206 
5207 	if (cnp->cn_pnbuf[0] == '\0') {
5208 		return (cache_fplookup_emptypath(fpl));
5209 	}
5210 
5211 	if (cnp->cn_nameptr[0] == '\0') {
5212 		if (fpl->tvp == NULL) {
5213 			return (cache_fplookup_degenerate(fpl));
5214 		}
5215 		return (cache_fplookup_trailingslash(fpl));
5216 	}
5217 
5218 	if (cnp->cn_nameiop != LOOKUP) {
5219 		fpl->tvp = NULL;
5220 		return (cache_fplookup_modifying(fpl));
5221 	}
5222 
5223 	/*
5224 	 * Only try to fill in the component if it is the last one,
5225 	 * otherwise not only there may be several to handle but the
5226 	 * walk may be complicated.
5227 	 */
5228 	if (!cache_fpl_islastcn(ndp)) {
5229 		return (cache_fpl_partial(fpl));
5230 	}
5231 
5232 	/*
5233 	 * Regular lookup nulifies the slash, which we don't do here.
5234 	 * Don't take chances with filesystem routines seeing it for
5235 	 * the last entry.
5236 	 */
5237 	if (cache_fpl_istrailingslash(fpl)) {
5238 		return (cache_fpl_partial(fpl));
5239 	}
5240 
5241 	/*
5242 	 * Secure access to dvp; check cache_fplookup_partial_setup for
5243 	 * reasoning.
5244 	 */
5245 	dvs = vget_prep_smr(dvp);
5246 	cache_fpl_smr_exit(fpl);
5247 	if (__predict_false(dvs == VGET_NONE)) {
5248 		return (cache_fpl_aborted(fpl));
5249 	}
5250 
5251 	vget_finish_ref(dvp, dvs);
5252 	if (!vn_seqc_consistent(dvp, dvp_seqc)) {
5253 		vrele(dvp);
5254 		return (cache_fpl_aborted(fpl));
5255 	}
5256 
5257 	error = vn_lock(dvp, LK_SHARED);
5258 	if (__predict_false(error != 0)) {
5259 		vrele(dvp);
5260 		return (cache_fpl_aborted(fpl));
5261 	}
5262 
5263 	tvp = NULL;
5264 	/*
5265 	 * TODO: provide variants which don't require locking either vnode.
5266 	 */
5267 	cnp->cn_flags |= ISLASTCN | MAKEENTRY;
5268 	cnp->cn_lkflags = LK_SHARED;
5269 	if ((cnp->cn_flags & LOCKSHARED) == 0) {
5270 		cnp->cn_lkflags = LK_EXCLUSIVE;
5271 	}
5272 	error = VOP_LOOKUP(dvp, &tvp, cnp);
5273 	switch (error) {
5274 	case EJUSTRETURN:
5275 	case 0:
5276 		break;
5277 	case ENOTDIR:
5278 	case ENOENT:
5279 		vput(dvp);
5280 		return (cache_fpl_handled_error(fpl, error));
5281 	default:
5282 		vput(dvp);
5283 		return (cache_fpl_aborted(fpl));
5284 	}
5285 
5286 	fpl->tvp = tvp;
5287 
5288 	if (tvp == NULL) {
5289 		MPASS(error == EJUSTRETURN);
5290 		if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) {
5291 			vput(dvp);
5292 		} else if ((cnp->cn_flags & LOCKPARENT) == 0) {
5293 			VOP_UNLOCK(dvp);
5294 		}
5295 		return (cache_fpl_handled(fpl));
5296 	}
5297 
5298 	if (tvp->v_type == VLNK) {
5299 		if ((cnp->cn_flags & FOLLOW) != 0) {
5300 			vput(dvp);
5301 			vput(tvp);
5302 			return (cache_fpl_aborted(fpl));
5303 		}
5304 	}
5305 
5306 	if (__predict_false(cache_fplookup_is_mp(fpl))) {
5307 		vput(dvp);
5308 		vput(tvp);
5309 		return (cache_fpl_aborted(fpl));
5310 	}
5311 
5312 	if ((cnp->cn_flags & LOCKLEAF) == 0) {
5313 		VOP_UNLOCK(tvp);
5314 	}
5315 
5316 	if ((cnp->cn_flags & (WANTPARENT | LOCKPARENT)) == 0) {
5317 		vput(dvp);
5318 	} else if ((cnp->cn_flags & LOCKPARENT) == 0) {
5319 		VOP_UNLOCK(dvp);
5320 	}
5321 	return (cache_fpl_handled(fpl));
5322 }
5323 
5324 static int __noinline
cache_fplookup_dot(struct cache_fpl * fpl)5325 cache_fplookup_dot(struct cache_fpl *fpl)
5326 {
5327 	int error;
5328 
5329 	MPASS(!seqc_in_modify(fpl->dvp_seqc));
5330 
5331 	if (__predict_false(fpl->dvp->v_type != VDIR)) {
5332 		cache_fpl_smr_exit(fpl);
5333 		return (cache_fpl_handled_error(fpl, ENOTDIR));
5334 	}
5335 
5336 	/*
5337 	 * Just re-assign the value. seqc will be checked later for the first
5338 	 * non-dot path component in line and/or before deciding to return the
5339 	 * vnode.
5340 	 */
5341 	fpl->tvp = fpl->dvp;
5342 	fpl->tvp_seqc = fpl->dvp_seqc;
5343 
5344 	SDT_PROBE3(vfs, namecache, lookup, hit, fpl->dvp, ".", fpl->dvp);
5345 
5346 	error = 0;
5347 	if (cache_fplookup_is_mp(fpl)) {
5348 		error = cache_fplookup_cross_mount(fpl);
5349 	}
5350 	return (error);
5351 }
5352 
5353 static int __noinline
cache_fplookup_dotdot(struct cache_fpl * fpl)5354 cache_fplookup_dotdot(struct cache_fpl *fpl)
5355 {
5356 	struct nameidata *ndp;
5357 	struct namecache *ncp;
5358 	struct vnode *dvp;
5359 	u_char nc_flag;
5360 
5361 	ndp = fpl->ndp;
5362 	dvp = fpl->dvp;
5363 
5364 	MPASS(cache_fpl_isdotdot(fpl->cnp));
5365 
5366 	/*
5367 	 * XXX this is racy the same way regular lookup is
5368 	 */
5369 	if (vfs_lookup_isroot(ndp, dvp)) {
5370 		fpl->tvp = dvp;
5371 		fpl->tvp_seqc = vn_seqc_read_any(dvp);
5372 		if (seqc_in_modify(fpl->tvp_seqc)) {
5373 			return (cache_fpl_aborted(fpl));
5374 		}
5375 		return (0);
5376 	}
5377 
5378 	if ((dvp->v_vflag & VV_ROOT) != 0) {
5379 		/*
5380 		 * TODO
5381 		 * The opposite of climb mount is needed here.
5382 		 */
5383 		return (cache_fpl_partial(fpl));
5384 	}
5385 
5386 	if (__predict_false(dvp->v_type != VDIR)) {
5387 		cache_fpl_smr_exit(fpl);
5388 		return (cache_fpl_handled_error(fpl, ENOTDIR));
5389 	}
5390 
5391 	ncp = atomic_load_consume_ptr(&dvp->v_cache_dd);
5392 	if (ncp == NULL) {
5393 		return (cache_fpl_aborted(fpl));
5394 	}
5395 
5396 	nc_flag = atomic_load_char(&ncp->nc_flag);
5397 	if ((nc_flag & NCF_ISDOTDOT) != 0) {
5398 		if ((nc_flag & NCF_NEGATIVE) != 0)
5399 			return (cache_fpl_aborted(fpl));
5400 		fpl->tvp = ncp->nc_vp;
5401 	} else {
5402 		fpl->tvp = ncp->nc_dvp;
5403 	}
5404 
5405 	fpl->tvp_seqc = vn_seqc_read_any(fpl->tvp);
5406 	if (seqc_in_modify(fpl->tvp_seqc)) {
5407 		return (cache_fpl_partial(fpl));
5408 	}
5409 
5410 	/*
5411 	 * Acquire fence provided by vn_seqc_read_any above.
5412 	 */
5413 	if (__predict_false(atomic_load_ptr(&dvp->v_cache_dd) != ncp)) {
5414 		return (cache_fpl_aborted(fpl));
5415 	}
5416 
5417 	if (!cache_ncp_canuse(ncp)) {
5418 		return (cache_fpl_aborted(fpl));
5419 	}
5420 
5421 	return (0);
5422 }
5423 
5424 static int __noinline
cache_fplookup_neg(struct cache_fpl * fpl,struct namecache * ncp,uint32_t hash)5425 cache_fplookup_neg(struct cache_fpl *fpl, struct namecache *ncp, uint32_t hash)
5426 {
5427 	u_char nc_flag __diagused;
5428 	bool neg_promote;
5429 
5430 #ifdef INVARIANTS
5431 	nc_flag = atomic_load_char(&ncp->nc_flag);
5432 	MPASS((nc_flag & NCF_NEGATIVE) != 0);
5433 #endif
5434 	/*
5435 	 * If they want to create an entry we need to replace this one.
5436 	 */
5437 	if (__predict_false(fpl->cnp->cn_nameiop != LOOKUP)) {
5438 		fpl->tvp = NULL;
5439 		return (cache_fplookup_modifying(fpl));
5440 	}
5441 	neg_promote = cache_neg_hit_prep(ncp);
5442 	if (!cache_fpl_neg_ncp_canuse(ncp)) {
5443 		cache_neg_hit_abort(ncp);
5444 		return (cache_fpl_partial(fpl));
5445 	}
5446 	if (neg_promote) {
5447 		return (cache_fplookup_negative_promote(fpl, ncp, hash));
5448 	}
5449 	cache_neg_hit_finish(ncp);
5450 	cache_fpl_smr_exit(fpl);
5451 	return (cache_fpl_handled_error(fpl, ENOENT));
5452 }
5453 
5454 /*
5455  * Resolve a symlink. Called by filesystem-specific routines.
5456  *
5457  * Code flow is:
5458  * ... -> cache_fplookup_symlink -> VOP_FPLOOKUP_SYMLINK -> cache_symlink_resolve
5459  */
5460 int
cache_symlink_resolve(struct cache_fpl * fpl,const char * string,size_t len)5461 cache_symlink_resolve(struct cache_fpl *fpl, const char *string, size_t len)
5462 {
5463 	struct nameidata *ndp;
5464 	struct componentname *cnp;
5465 	size_t adjust;
5466 
5467 	ndp = fpl->ndp;
5468 	cnp = fpl->cnp;
5469 
5470 	if (__predict_false(len == 0)) {
5471 		return (ENOENT);
5472 	}
5473 
5474 	if (__predict_false(len > MAXPATHLEN - 2)) {
5475 		if (cache_fpl_istrailingslash(fpl)) {
5476 			return (EAGAIN);
5477 		}
5478 	}
5479 
5480 	ndp->ni_pathlen = fpl->nulchar - cnp->cn_nameptr - cnp->cn_namelen + 1;
5481 #ifdef INVARIANTS
5482 	if (ndp->ni_pathlen != fpl->debug.ni_pathlen) {
5483 		panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n",
5484 		    __func__, ndp->ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar,
5485 		    cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf);
5486 	}
5487 #endif
5488 
5489 	if (__predict_false(len + ndp->ni_pathlen > MAXPATHLEN)) {
5490 		return (ENAMETOOLONG);
5491 	}
5492 
5493 	if (__predict_false(ndp->ni_loopcnt++ >= MAXSYMLINKS)) {
5494 		return (ELOOP);
5495 	}
5496 
5497 	adjust = len;
5498 	if (ndp->ni_pathlen > 1) {
5499 		bcopy(ndp->ni_next, cnp->cn_pnbuf + len, ndp->ni_pathlen);
5500 	} else {
5501 		if (cache_fpl_istrailingslash(fpl)) {
5502 			adjust = len + 1;
5503 			cnp->cn_pnbuf[len] = '/';
5504 			cnp->cn_pnbuf[len + 1] = '\0';
5505 		} else {
5506 			cnp->cn_pnbuf[len] = '\0';
5507 		}
5508 	}
5509 	bcopy(string, cnp->cn_pnbuf, len);
5510 
5511 	ndp->ni_pathlen += adjust;
5512 	cache_fpl_pathlen_add(fpl, adjust);
5513 	cnp->cn_nameptr = cnp->cn_pnbuf;
5514 	fpl->nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1];
5515 	fpl->tvp = NULL;
5516 	return (0);
5517 }
5518 
5519 static int __noinline
cache_fplookup_symlink(struct cache_fpl * fpl)5520 cache_fplookup_symlink(struct cache_fpl *fpl)
5521 {
5522 	struct mount *mp;
5523 	struct nameidata *ndp;
5524 	struct componentname *cnp;
5525 	struct vnode *dvp, *tvp;
5526 	struct pwd *pwd;
5527 	int error;
5528 
5529 	ndp = fpl->ndp;
5530 	cnp = fpl->cnp;
5531 	dvp = fpl->dvp;
5532 	tvp = fpl->tvp;
5533 	pwd = *(fpl->pwd);
5534 
5535 	if (cache_fpl_islastcn(ndp)) {
5536 		if ((cnp->cn_flags & FOLLOW) == 0) {
5537 			return (cache_fplookup_final(fpl));
5538 		}
5539 	}
5540 
5541 	mp = atomic_load_ptr(&dvp->v_mount);
5542 	if (__predict_false(mp == NULL)) {
5543 		return (cache_fpl_aborted(fpl));
5544 	}
5545 
5546 	/*
5547 	 * Note this check races against setting the flag just like regular
5548 	 * lookup.
5549 	 */
5550 	if (__predict_false((mp->mnt_flag & MNT_NOSYMFOLLOW) != 0)) {
5551 		cache_fpl_smr_exit(fpl);
5552 		return (cache_fpl_handled_error(fpl, EACCES));
5553 	}
5554 
5555 	error = VOP_FPLOOKUP_SYMLINK(tvp, fpl);
5556 	if (__predict_false(error != 0)) {
5557 		switch (error) {
5558 		case EAGAIN:
5559 			return (cache_fpl_partial(fpl));
5560 		case ENOENT:
5561 		case ENAMETOOLONG:
5562 		case ELOOP:
5563 			cache_fpl_smr_exit(fpl);
5564 			return (cache_fpl_handled_error(fpl, error));
5565 		default:
5566 			return (cache_fpl_aborted(fpl));
5567 		}
5568 	}
5569 
5570 	if (*(cnp->cn_nameptr) == '/') {
5571 		fpl->dvp = cache_fpl_handle_root(fpl);
5572 		fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp);
5573 		if (seqc_in_modify(fpl->dvp_seqc)) {
5574 			return (cache_fpl_aborted(fpl));
5575 		}
5576 		/*
5577 		 * The main loop assumes that ->dvp points to a vnode belonging
5578 		 * to a filesystem which can do lockless lookup, but the absolute
5579 		 * symlink can be wandering off to one which does not.
5580 		 */
5581 		mp = atomic_load_ptr(&fpl->dvp->v_mount);
5582 		if (__predict_false(mp == NULL)) {
5583 			return (cache_fpl_aborted(fpl));
5584 		}
5585 		if (!cache_fplookup_mp_supported(mp)) {
5586 			cache_fpl_checkpoint(fpl);
5587 			return (cache_fpl_partial(fpl));
5588 		}
5589 		if (__predict_false(pwd->pwd_adir != pwd->pwd_rdir)) {
5590 			return (cache_fpl_aborted(fpl));
5591 		}
5592 	}
5593 	return (0);
5594 }
5595 
5596 static int
cache_fplookup_next(struct cache_fpl * fpl)5597 cache_fplookup_next(struct cache_fpl *fpl)
5598 {
5599 	struct componentname *cnp;
5600 	struct namecache *ncp;
5601 	struct vnode *dvp, *tvp;
5602 	u_char nc_flag;
5603 	uint32_t hash;
5604 	int error;
5605 
5606 	cnp = fpl->cnp;
5607 	dvp = fpl->dvp;
5608 	hash = fpl->hash;
5609 
5610 	if (__predict_false(cnp->cn_nameptr[0] == '.')) {
5611 		if (cnp->cn_namelen == 1) {
5612 			return (cache_fplookup_dot(fpl));
5613 		}
5614 		if (cnp->cn_namelen == 2 && cnp->cn_nameptr[1] == '.') {
5615 			return (cache_fplookup_dotdot(fpl));
5616 		}
5617 	}
5618 
5619 	MPASS(!cache_fpl_isdotdot(cnp));
5620 
5621 	ncp = cache_ncp_find(dvp, cnp, hash);
5622 	if (__predict_false(ncp == NULL)) {
5623 		return (cache_fplookup_noentry(fpl));
5624 	}
5625 
5626 	tvp = atomic_load_ptr(&ncp->nc_vp);
5627 	nc_flag = atomic_load_char(&ncp->nc_flag);
5628 	if ((nc_flag & NCF_NEGATIVE) != 0) {
5629 		return (cache_fplookup_neg(fpl, ncp, hash));
5630 	}
5631 
5632 	if (!cache_ncp_canuse(ncp)) {
5633 		return (cache_fpl_partial(fpl));
5634 	}
5635 
5636 	fpl->tvp = tvp;
5637 	fpl->tvp_seqc = vn_seqc_read_any(tvp);
5638 	if (seqc_in_modify(fpl->tvp_seqc)) {
5639 		return (cache_fpl_partial(fpl));
5640 	}
5641 
5642 	counter_u64_add(numposhits, 1);
5643 	SDT_PROBE3(vfs, namecache, lookup, hit, dvp, ncp->nc_name, tvp);
5644 
5645 	error = 0;
5646 	if (cache_fplookup_is_mp(fpl)) {
5647 		error = cache_fplookup_cross_mount(fpl);
5648 	}
5649 	return (error);
5650 }
5651 
5652 static bool
cache_fplookup_mp_supported(struct mount * mp)5653 cache_fplookup_mp_supported(struct mount *mp)
5654 {
5655 
5656 	MPASS(mp != NULL);
5657 	if ((mp->mnt_kern_flag & MNTK_FPLOOKUP) == 0)
5658 		return (false);
5659 	return (true);
5660 }
5661 
5662 /*
5663  * Walk up the mount stack (if any).
5664  *
5665  * Correctness is provided in the following ways:
5666  * - all vnodes are protected from freeing with SMR
5667  * - struct mount objects are type stable making them always safe to access
5668  * - stability of the particular mount is provided by busying it
5669  * - relationship between the vnode which is mounted on and the mount is
5670  *   verified with the vnode sequence counter after busying
5671  * - association between root vnode of the mount and the mount is protected
5672  *   by busy
5673  *
5674  * From that point on we can read the sequence counter of the root vnode
5675  * and get the next mount on the stack (if any) using the same protection.
5676  *
5677  * By the end of successful walk we are guaranteed the reached state was
5678  * indeed present at least at some point which matches the regular lookup.
5679  */
5680 static int __noinline
cache_fplookup_climb_mount(struct cache_fpl * fpl)5681 cache_fplookup_climb_mount(struct cache_fpl *fpl)
5682 {
5683 	struct mount *mp, *prev_mp;
5684 	struct mount_pcpu *mpcpu, *prev_mpcpu;
5685 	struct vnode *vp;
5686 	seqc_t vp_seqc;
5687 
5688 	vp = fpl->tvp;
5689 	vp_seqc = fpl->tvp_seqc;
5690 
5691 	VNPASS(vp->v_type == VDIR || vp->v_type == VREG || vp->v_type == VBAD, vp);
5692 	mp = atomic_load_ptr(&vp->v_mountedhere);
5693 	if (__predict_false(mp == NULL)) {
5694 		return (0);
5695 	}
5696 
5697 	prev_mp = NULL;
5698 	for (;;) {
5699 		if (!vfs_op_thread_enter_crit(mp, mpcpu)) {
5700 			if (prev_mp != NULL)
5701 				vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
5702 			return (cache_fpl_partial(fpl));
5703 		}
5704 		if (prev_mp != NULL)
5705 			vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
5706 		if (!vn_seqc_consistent(vp, vp_seqc)) {
5707 			vfs_op_thread_exit_crit(mp, mpcpu);
5708 			return (cache_fpl_partial(fpl));
5709 		}
5710 		if (!cache_fplookup_mp_supported(mp)) {
5711 			vfs_op_thread_exit_crit(mp, mpcpu);
5712 			return (cache_fpl_partial(fpl));
5713 		}
5714 		vp = atomic_load_ptr(&mp->mnt_rootvnode);
5715 		if (vp == NULL) {
5716 			vfs_op_thread_exit_crit(mp, mpcpu);
5717 			return (cache_fpl_partial(fpl));
5718 		}
5719 		vp_seqc = vn_seqc_read_any(vp);
5720 		if (seqc_in_modify(vp_seqc)) {
5721 			vfs_op_thread_exit_crit(mp, mpcpu);
5722 			return (cache_fpl_partial(fpl));
5723 		}
5724 		prev_mp = mp;
5725 		prev_mpcpu = mpcpu;
5726 		mp = atomic_load_ptr(&vp->v_mountedhere);
5727 		if (mp == NULL)
5728 			break;
5729 	}
5730 
5731 	vfs_op_thread_exit_crit(prev_mp, prev_mpcpu);
5732 	fpl->tvp = vp;
5733 	fpl->tvp_seqc = vp_seqc;
5734 	return (0);
5735 }
5736 
5737 static int __noinline
cache_fplookup_cross_mount(struct cache_fpl * fpl)5738 cache_fplookup_cross_mount(struct cache_fpl *fpl)
5739 {
5740 	struct mount *mp;
5741 	struct mount_pcpu *mpcpu;
5742 	struct vnode *vp;
5743 	seqc_t vp_seqc;
5744 
5745 	vp = fpl->tvp;
5746 	vp_seqc = fpl->tvp_seqc;
5747 
5748 	VNPASS(vp->v_type == VDIR || vp->v_type == VREG || vp->v_type == VBAD, vp);
5749 	mp = atomic_load_ptr(&vp->v_mountedhere);
5750 	if (__predict_false(mp == NULL)) {
5751 		return (0);
5752 	}
5753 
5754 	if (!vfs_op_thread_enter_crit(mp, mpcpu)) {
5755 		return (cache_fpl_partial(fpl));
5756 	}
5757 	if (!vn_seqc_consistent(vp, vp_seqc)) {
5758 		vfs_op_thread_exit_crit(mp, mpcpu);
5759 		return (cache_fpl_partial(fpl));
5760 	}
5761 	if (!cache_fplookup_mp_supported(mp)) {
5762 		vfs_op_thread_exit_crit(mp, mpcpu);
5763 		return (cache_fpl_partial(fpl));
5764 	}
5765 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
5766 	if (__predict_false(vp == NULL)) {
5767 		vfs_op_thread_exit_crit(mp, mpcpu);
5768 		return (cache_fpl_partial(fpl));
5769 	}
5770 	vp_seqc = vn_seqc_read_any(vp);
5771 	vfs_op_thread_exit_crit(mp, mpcpu);
5772 	if (seqc_in_modify(vp_seqc)) {
5773 		return (cache_fpl_partial(fpl));
5774 	}
5775 	mp = atomic_load_ptr(&vp->v_mountedhere);
5776 	if (__predict_false(mp != NULL)) {
5777 		/*
5778 		 * There are possibly more mount points on top.
5779 		 * Normally this does not happen so for simplicity just start
5780 		 * over.
5781 		 */
5782 		return (cache_fplookup_climb_mount(fpl));
5783 	}
5784 
5785 	fpl->tvp = vp;
5786 	fpl->tvp_seqc = vp_seqc;
5787 	return (0);
5788 }
5789 
5790 /*
5791  * Check if a vnode is mounted on.
5792  */
5793 static bool
cache_fplookup_is_mp(struct cache_fpl * fpl)5794 cache_fplookup_is_mp(struct cache_fpl *fpl)
5795 {
5796 	struct vnode *vp;
5797 
5798 	vp = fpl->tvp;
5799 	return ((vn_irflag_read(vp) & VIRF_MOUNTPOINT) != 0);
5800 }
5801 
5802 /*
5803  * Parse the path.
5804  *
5805  * The code was originally copy-pasted from regular lookup and despite
5806  * clean ups leaves performance on the table. Any modifications here
5807  * must take into account that in case off fallback the resulting
5808  * nameidata state has to be compatible with the original.
5809  */
5810 
5811 /*
5812  * Debug ni_pathlen tracking.
5813  */
5814 #ifdef INVARIANTS
5815 static void
cache_fpl_pathlen_add(struct cache_fpl * fpl,size_t n)5816 cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n)
5817 {
5818 
5819 	fpl->debug.ni_pathlen += n;
5820 	KASSERT(fpl->debug.ni_pathlen <= PATH_MAX,
5821 	    ("%s: pathlen overflow to %zd\n", __func__, fpl->debug.ni_pathlen));
5822 }
5823 
5824 static void
cache_fpl_pathlen_sub(struct cache_fpl * fpl,size_t n)5825 cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n)
5826 {
5827 
5828 	fpl->debug.ni_pathlen -= n;
5829 	KASSERT(fpl->debug.ni_pathlen <= PATH_MAX,
5830 	    ("%s: pathlen underflow to %zd\n", __func__, fpl->debug.ni_pathlen));
5831 }
5832 
5833 static void
cache_fpl_pathlen_inc(struct cache_fpl * fpl)5834 cache_fpl_pathlen_inc(struct cache_fpl *fpl)
5835 {
5836 
5837 	cache_fpl_pathlen_add(fpl, 1);
5838 }
5839 
5840 static void
cache_fpl_pathlen_dec(struct cache_fpl * fpl)5841 cache_fpl_pathlen_dec(struct cache_fpl *fpl)
5842 {
5843 
5844 	cache_fpl_pathlen_sub(fpl, 1);
5845 }
5846 #else
5847 static void
cache_fpl_pathlen_add(struct cache_fpl * fpl,size_t n)5848 cache_fpl_pathlen_add(struct cache_fpl *fpl, size_t n)
5849 {
5850 }
5851 
5852 static void
cache_fpl_pathlen_sub(struct cache_fpl * fpl,size_t n)5853 cache_fpl_pathlen_sub(struct cache_fpl *fpl, size_t n)
5854 {
5855 }
5856 
5857 static void
cache_fpl_pathlen_inc(struct cache_fpl * fpl)5858 cache_fpl_pathlen_inc(struct cache_fpl *fpl)
5859 {
5860 }
5861 
5862 static void
cache_fpl_pathlen_dec(struct cache_fpl * fpl)5863 cache_fpl_pathlen_dec(struct cache_fpl *fpl)
5864 {
5865 }
5866 #endif
5867 
5868 static void
cache_fplookup_parse(struct cache_fpl * fpl)5869 cache_fplookup_parse(struct cache_fpl *fpl)
5870 {
5871 	struct nameidata *ndp;
5872 	struct componentname *cnp;
5873 	struct vnode *dvp;
5874 	char *cp;
5875 	uint32_t hash;
5876 
5877 	ndp = fpl->ndp;
5878 	cnp = fpl->cnp;
5879 	dvp = fpl->dvp;
5880 
5881 	/*
5882 	 * Find the end of this path component, it is either / or nul.
5883 	 *
5884 	 * Store / as a temporary sentinel so that we only have one character
5885 	 * to test for. Pathnames tend to be short so this should not be
5886 	 * resulting in cache misses.
5887 	 *
5888 	 * TODO: fix this to be word-sized.
5889 	 */
5890 	MPASS(&cnp->cn_nameptr[fpl->debug.ni_pathlen - 1] >= cnp->cn_pnbuf);
5891 	KASSERT(&cnp->cn_nameptr[fpl->debug.ni_pathlen - 1] == fpl->nulchar,
5892 	    ("%s: mismatch between pathlen (%zu) and nulchar (%p != %p), string [%s]\n",
5893 	    __func__, fpl->debug.ni_pathlen, &cnp->cn_nameptr[fpl->debug.ni_pathlen - 1],
5894 	    fpl->nulchar, cnp->cn_pnbuf));
5895 	KASSERT(*fpl->nulchar == '\0',
5896 	    ("%s: expected nul at %p; string [%s]\n", __func__, fpl->nulchar,
5897 	    cnp->cn_pnbuf));
5898 	hash = cache_get_hash_iter_start(dvp);
5899 	*fpl->nulchar = '/';
5900 	for (cp = cnp->cn_nameptr; *cp != '/'; cp++) {
5901 		KASSERT(*cp != '\0',
5902 		    ("%s: encountered unexpected nul; string [%s]\n", __func__,
5903 		    cnp->cn_nameptr));
5904 		hash = cache_get_hash_iter(*cp, hash);
5905 		continue;
5906 	}
5907 	*fpl->nulchar = '\0';
5908 	fpl->hash = cache_get_hash_iter_finish(hash);
5909 
5910 	cnp->cn_namelen = cp - cnp->cn_nameptr;
5911 	cache_fpl_pathlen_sub(fpl, cnp->cn_namelen);
5912 
5913 #ifdef INVARIANTS
5914 	/*
5915 	 * cache_get_hash only accepts lengths up to NAME_MAX. This is fine since
5916 	 * we are going to fail this lookup with ENAMETOOLONG (see below).
5917 	 */
5918 	if (cnp->cn_namelen <= NAME_MAX) {
5919 		if (fpl->hash != cache_get_hash(cnp->cn_nameptr, cnp->cn_namelen, dvp)) {
5920 			panic("%s: mismatched hash for [%s] len %ld", __func__,
5921 			    cnp->cn_nameptr, cnp->cn_namelen);
5922 		}
5923 	}
5924 #endif
5925 
5926 	/*
5927 	 * Hack: we have to check if the found path component's length exceeds
5928 	 * NAME_MAX. However, the condition is very rarely true and check can
5929 	 * be elided in the common case -- if an entry was found in the cache,
5930 	 * then it could not have been too long to begin with.
5931 	 */
5932 	ndp->ni_next = cp;
5933 }
5934 
5935 static void
cache_fplookup_parse_advance(struct cache_fpl * fpl)5936 cache_fplookup_parse_advance(struct cache_fpl *fpl)
5937 {
5938 	struct nameidata *ndp;
5939 	struct componentname *cnp;
5940 
5941 	ndp = fpl->ndp;
5942 	cnp = fpl->cnp;
5943 
5944 	cnp->cn_nameptr = ndp->ni_next;
5945 	KASSERT(*(cnp->cn_nameptr) == '/',
5946 	    ("%s: should have seen slash at %p ; buf %p [%s]\n", __func__,
5947 	    cnp->cn_nameptr, cnp->cn_pnbuf, cnp->cn_pnbuf));
5948 	cnp->cn_nameptr++;
5949 	cache_fpl_pathlen_dec(fpl);
5950 }
5951 
5952 /*
5953  * Skip spurious slashes in a pathname (e.g., "foo///bar") and retry.
5954  *
5955  * Lockless lookup tries to elide checking for spurious slashes and should they
5956  * be present is guaranteed to fail to find an entry. In this case the caller
5957  * must check if the name starts with a slash and call this routine.  It is
5958  * going to fast forward across the spurious slashes and set the state up for
5959  * retry.
5960  */
5961 static int __noinline
cache_fplookup_skip_slashes(struct cache_fpl * fpl)5962 cache_fplookup_skip_slashes(struct cache_fpl *fpl)
5963 {
5964 	struct nameidata *ndp;
5965 	struct componentname *cnp;
5966 
5967 	ndp = fpl->ndp;
5968 	cnp = fpl->cnp;
5969 
5970 	MPASS(*(cnp->cn_nameptr) == '/');
5971 	do {
5972 		cnp->cn_nameptr++;
5973 		cache_fpl_pathlen_dec(fpl);
5974 	} while (*(cnp->cn_nameptr) == '/');
5975 
5976 	/*
5977 	 * Go back to one slash so that cache_fplookup_parse_advance has
5978 	 * something to skip.
5979 	 */
5980 	cnp->cn_nameptr--;
5981 	cache_fpl_pathlen_inc(fpl);
5982 
5983 	/*
5984 	 * cache_fplookup_parse_advance starts from ndp->ni_next
5985 	 */
5986 	ndp->ni_next = cnp->cn_nameptr;
5987 
5988 	/*
5989 	 * See cache_fplookup_dot.
5990 	 */
5991 	fpl->tvp = fpl->dvp;
5992 	fpl->tvp_seqc = fpl->dvp_seqc;
5993 
5994 	return (0);
5995 }
5996 
5997 /*
5998  * Handle trailing slashes (e.g., "foo/").
5999  *
6000  * If a trailing slash is found the terminal vnode must be a directory.
6001  * Regular lookup shortens the path by nulifying the first trailing slash and
6002  * sets the TRAILINGSLASH flag to denote this took place. There are several
6003  * checks on it performed later.
6004  *
6005  * Similarly to spurious slashes, lockless lookup handles this in a speculative
6006  * manner relying on an invariant that a non-directory vnode will get a miss.
6007  * In this case cn_nameptr[0] == '\0' and cn_namelen == 0.
6008  *
6009  * Thus for a path like "foo/bar/" the code unwinds the state back to "bar/"
6010  * and denotes this is the last path component, which avoids looping back.
6011  *
6012  * Only plain lookups are supported for now to restrict corner cases to handle.
6013  */
6014 static int __noinline
cache_fplookup_trailingslash(struct cache_fpl * fpl)6015 cache_fplookup_trailingslash(struct cache_fpl *fpl)
6016 {
6017 #ifdef INVARIANTS
6018 	size_t ni_pathlen;
6019 #endif
6020 	struct nameidata *ndp;
6021 	struct componentname *cnp;
6022 	struct namecache *ncp;
6023 	struct vnode *tvp;
6024 	char *cn_nameptr_orig, *cn_nameptr_slash;
6025 	seqc_t tvp_seqc;
6026 	u_char nc_flag;
6027 
6028 	ndp = fpl->ndp;
6029 	cnp = fpl->cnp;
6030 	tvp = fpl->tvp;
6031 	tvp_seqc = fpl->tvp_seqc;
6032 
6033 	MPASS(fpl->dvp == fpl->tvp);
6034 	KASSERT(cache_fpl_istrailingslash(fpl),
6035 	    ("%s: expected trailing slash at %p; string [%s]\n", __func__, fpl->nulchar - 1,
6036 	    cnp->cn_pnbuf));
6037 	KASSERT(cnp->cn_nameptr[0] == '\0',
6038 	    ("%s: expected nul char at %p; string [%s]\n", __func__, &cnp->cn_nameptr[0],
6039 	    cnp->cn_pnbuf));
6040 	KASSERT(cnp->cn_namelen == 0,
6041 	    ("%s: namelen 0 but got %ld; string [%s]\n", __func__, cnp->cn_namelen,
6042 	    cnp->cn_pnbuf));
6043 	MPASS(cnp->cn_nameptr > cnp->cn_pnbuf);
6044 
6045 	if (cnp->cn_nameiop != LOOKUP) {
6046 		return (cache_fpl_aborted(fpl));
6047 	}
6048 
6049 	if (__predict_false(tvp->v_type != VDIR)) {
6050 		if (!vn_seqc_consistent(tvp, tvp_seqc)) {
6051 			return (cache_fpl_aborted(fpl));
6052 		}
6053 		cache_fpl_smr_exit(fpl);
6054 		return (cache_fpl_handled_error(fpl, ENOTDIR));
6055 	}
6056 
6057 	/*
6058 	 * Denote the last component.
6059 	 */
6060 	ndp->ni_next = &cnp->cn_nameptr[0];
6061 	MPASS(cache_fpl_islastcn(ndp));
6062 
6063 	/*
6064 	 * Unwind trailing slashes.
6065 	 */
6066 	cn_nameptr_orig = cnp->cn_nameptr;
6067 	while (cnp->cn_nameptr >= cnp->cn_pnbuf) {
6068 		cnp->cn_nameptr--;
6069 		if (cnp->cn_nameptr[0] != '/') {
6070 			break;
6071 		}
6072 	}
6073 
6074 	/*
6075 	 * Unwind to the beginning of the path component.
6076 	 *
6077 	 * Note the path may or may not have started with a slash.
6078 	 */
6079 	cn_nameptr_slash = cnp->cn_nameptr;
6080 	while (cnp->cn_nameptr > cnp->cn_pnbuf) {
6081 		cnp->cn_nameptr--;
6082 		if (cnp->cn_nameptr[0] == '/') {
6083 			break;
6084 		}
6085 	}
6086 	if (cnp->cn_nameptr[0] == '/') {
6087 		cnp->cn_nameptr++;
6088 	}
6089 
6090 	cnp->cn_namelen = cn_nameptr_slash - cnp->cn_nameptr + 1;
6091 	cache_fpl_pathlen_add(fpl, cn_nameptr_orig - cnp->cn_nameptr);
6092 	cache_fpl_checkpoint(fpl);
6093 
6094 #ifdef INVARIANTS
6095 	ni_pathlen = fpl->nulchar - cnp->cn_nameptr + 1;
6096 	if (ni_pathlen != fpl->debug.ni_pathlen) {
6097 		panic("%s: mismatch (%zu != %zu) nulchar %p nameptr %p [%s] ; full string [%s]\n",
6098 		    __func__, ni_pathlen, fpl->debug.ni_pathlen, fpl->nulchar,
6099 		    cnp->cn_nameptr, cnp->cn_nameptr, cnp->cn_pnbuf);
6100 	}
6101 #endif
6102 
6103 	/*
6104 	 * If this was a "./" lookup the parent directory is already correct.
6105 	 */
6106 	if (cnp->cn_nameptr[0] == '.' && cnp->cn_namelen == 1) {
6107 		return (0);
6108 	}
6109 
6110 	/*
6111 	 * Otherwise we need to look it up.
6112 	 */
6113 	tvp = fpl->tvp;
6114 	ncp = atomic_load_consume_ptr(&tvp->v_cache_dd);
6115 	if (__predict_false(ncp == NULL)) {
6116 		return (cache_fpl_aborted(fpl));
6117 	}
6118 	nc_flag = atomic_load_char(&ncp->nc_flag);
6119 	if ((nc_flag & NCF_ISDOTDOT) != 0) {
6120 		return (cache_fpl_aborted(fpl));
6121 	}
6122 	fpl->dvp = ncp->nc_dvp;
6123 	fpl->dvp_seqc = vn_seqc_read_any(fpl->dvp);
6124 	if (seqc_in_modify(fpl->dvp_seqc)) {
6125 		return (cache_fpl_aborted(fpl));
6126 	}
6127 	return (0);
6128 }
6129 
6130 /*
6131  * See the API contract for VOP_FPLOOKUP_VEXEC.
6132  */
6133 static int __noinline
cache_fplookup_failed_vexec(struct cache_fpl * fpl,int error)6134 cache_fplookup_failed_vexec(struct cache_fpl *fpl, int error)
6135 {
6136 	struct componentname *cnp;
6137 	struct vnode *dvp;
6138 	seqc_t dvp_seqc;
6139 
6140 	cnp = fpl->cnp;
6141 	dvp = fpl->dvp;
6142 	dvp_seqc = fpl->dvp_seqc;
6143 
6144 	/*
6145 	 * Hack: delayed empty path checking.
6146 	 */
6147 	if (cnp->cn_pnbuf[0] == '\0') {
6148 		return (cache_fplookup_emptypath(fpl));
6149 	}
6150 
6151 	/*
6152 	 * TODO: Due to ignoring trailing slashes lookup will perform a
6153 	 * permission check on the last dir when it should not be doing it.  It
6154 	 * may fail, but said failure should be ignored. It is possible to fix
6155 	 * it up fully without resorting to regular lookup, but for now just
6156 	 * abort.
6157 	 */
6158 	if (cache_fpl_istrailingslash(fpl)) {
6159 		return (cache_fpl_aborted(fpl));
6160 	}
6161 
6162 	/*
6163 	 * Hack: delayed degenerate path checking.
6164 	 */
6165 	if (cnp->cn_nameptr[0] == '\0' && fpl->tvp == NULL) {
6166 		return (cache_fplookup_degenerate(fpl));
6167 	}
6168 
6169 	/*
6170 	 * Hack: delayed name len checking.
6171 	 */
6172 	if (__predict_false(cnp->cn_namelen > NAME_MAX)) {
6173 		cache_fpl_smr_exit(fpl);
6174 		return (cache_fpl_handled_error(fpl, ENAMETOOLONG));
6175 	}
6176 
6177 	/*
6178 	 * Hack: they may be looking up foo/bar, where foo is not a directory.
6179 	 * In such a case we need to return ENOTDIR, but we may happen to get
6180 	 * here with a different error.
6181 	 */
6182 	if (dvp->v_type != VDIR) {
6183 		error = ENOTDIR;
6184 	}
6185 
6186 	/*
6187 	 * Hack: handle O_SEARCH.
6188 	 *
6189 	 * Open Group Base Specifications Issue 7, 2018 edition states:
6190 	 * <quote>
6191 	 * If the access mode of the open file description associated with the
6192 	 * file descriptor is not O_SEARCH, the function shall check whether
6193 	 * directory searches are permitted using the current permissions of
6194 	 * the directory underlying the file descriptor. If the access mode is
6195 	 * O_SEARCH, the function shall not perform the check.
6196 	 * </quote>
6197 	 *
6198 	 * Regular lookup tests for the NOEXECCHECK flag for every path
6199 	 * component to decide whether to do the permission check. However,
6200 	 * since most lookups never have the flag (and when they do it is only
6201 	 * present for the first path component), lockless lookup only acts on
6202 	 * it if there is a permission problem. Here the flag is represented
6203 	 * with a boolean so that we don't have to clear it on the way out.
6204 	 *
6205 	 * For simplicity this always aborts.
6206 	 * TODO: check if this is the first lookup and ignore the permission
6207 	 * problem. Note the flag has to survive fallback (if it happens to be
6208 	 * performed).
6209 	 */
6210 	if (fpl->fsearch) {
6211 		return (cache_fpl_aborted(fpl));
6212 	}
6213 
6214 	switch (error) {
6215 	case EAGAIN:
6216 		if (!vn_seqc_consistent(dvp, dvp_seqc)) {
6217 			error = cache_fpl_aborted(fpl);
6218 		} else {
6219 			cache_fpl_partial(fpl);
6220 		}
6221 		break;
6222 	default:
6223 		if (!vn_seqc_consistent(dvp, dvp_seqc)) {
6224 			error = cache_fpl_aborted(fpl);
6225 		} else {
6226 			cache_fpl_smr_exit(fpl);
6227 			cache_fpl_handled_error(fpl, error);
6228 		}
6229 		break;
6230 	}
6231 	return (error);
6232 }
6233 
6234 static int
cache_fplookup_impl(struct vnode * dvp,struct cache_fpl * fpl)6235 cache_fplookup_impl(struct vnode *dvp, struct cache_fpl *fpl)
6236 {
6237 	struct nameidata *ndp;
6238 	struct componentname *cnp;
6239 	struct mount *mp;
6240 	int error;
6241 
6242 	ndp = fpl->ndp;
6243 	cnp = fpl->cnp;
6244 
6245 	cache_fpl_checkpoint(fpl);
6246 
6247 	/*
6248 	 * The vnode at hand is almost always stable, skip checking for it.
6249 	 * Worst case this postpones the check towards the end of the iteration
6250 	 * of the main loop.
6251 	 */
6252 	fpl->dvp = dvp;
6253 	fpl->dvp_seqc = vn_seqc_read_notmodify(fpl->dvp);
6254 
6255 	mp = atomic_load_ptr(&dvp->v_mount);
6256 	if (__predict_false(mp == NULL || !cache_fplookup_mp_supported(mp))) {
6257 		return (cache_fpl_aborted(fpl));
6258 	}
6259 
6260 	MPASS(fpl->tvp == NULL);
6261 
6262 	for (;;) {
6263 		cache_fplookup_parse(fpl);
6264 
6265 		error = VOP_FPLOOKUP_VEXEC(fpl->dvp, cnp->cn_cred);
6266 		if (__predict_false(error != 0)) {
6267 			error = cache_fplookup_failed_vexec(fpl, error);
6268 			break;
6269 		}
6270 
6271 		error = cache_fplookup_next(fpl);
6272 		if (__predict_false(cache_fpl_terminated(fpl))) {
6273 			break;
6274 		}
6275 
6276 		VNPASS(!seqc_in_modify(fpl->tvp_seqc), fpl->tvp);
6277 
6278 		if (fpl->tvp->v_type == VLNK) {
6279 			error = cache_fplookup_symlink(fpl);
6280 			if (cache_fpl_terminated(fpl)) {
6281 				break;
6282 			}
6283 		} else {
6284 			if (cache_fpl_islastcn(ndp)) {
6285 				error = cache_fplookup_final(fpl);
6286 				break;
6287 			}
6288 
6289 			if (!vn_seqc_consistent(fpl->dvp, fpl->dvp_seqc)) {
6290 				error = cache_fpl_aborted(fpl);
6291 				break;
6292 			}
6293 
6294 			fpl->dvp = fpl->tvp;
6295 			fpl->dvp_seqc = fpl->tvp_seqc;
6296 			cache_fplookup_parse_advance(fpl);
6297 		}
6298 
6299 		cache_fpl_checkpoint(fpl);
6300 	}
6301 
6302 	return (error);
6303 }
6304 
6305 /*
6306  * Fast path lookup protected with SMR and sequence counters.
6307  *
6308  * Note: all VOP_FPLOOKUP_VEXEC routines have a comment referencing this one.
6309  *
6310  * Filesystems can opt in by setting the MNTK_FPLOOKUP flag and meeting criteria
6311  * outlined at the end.
6312  *
6313  * Traversing from one vnode to another requires atomicity with regard to
6314  * permissions, mount points and of course their relative placement (if you are
6315  * looking up "bar" in "foo" and you found it, it better be in that directory
6316  * at the time).
6317  *
6318  * Normally this is accomplished with locking, but it comes with a significant
6319  * performance hit and is untenable as a fast path even in a moderate core
6320  * count environment (at the time of writing this comment this would be a
6321  * little south of 100).
6322  *
6323  * The same guarantee can be provided with a combination of safe memory
6324  * reclamation and sequence counters instead. If all operations which affect
6325  * the relationship between the current vnode and the one we are looking for
6326  * also modify the counter, we can verify whether all the conditions held as
6327  * we made the jump.
6328  *
6329  * See places which issue vn_seqc_write_begin()/vn_seqc_write_end() for
6330  * operations affected.
6331  *
6332  * Suppose the variable "cnp" contains lookup metadata (the path etc.), then
6333  * locked lookup conceptually looks like this:
6334  *
6335  * // lock the current directory
6336  * vn_lock(dvp);
6337  * for (;;) {
6338  *      // permission check
6339  * 	if (!canlookup(dvp, cnp))
6340  * 	    abort();
6341  * 	// look for the target name inside dvp
6342  *	tvp = findnext(dvp, cnp);
6343  *	vn_lock(tvp);
6344  *	// tvp is still guaranteed to be inside of dvp because of the lock on dvp
6345  *	vn_unlock(dvp);
6346  *      // dvp is unlocked. its state is now arbitrary, but that's fine as we
6347  *      // made the jump while everything relevant was correct, continue with tvp
6348  *      // as the directory to look up names in
6349  *	tvp = dvp;
6350  *	if (last)
6351  *	    break;
6352  *	// if not last loop back and continue until done
6353  * }
6354  * vget(tvp);
6355  * return (tvp);
6356  *
6357  * Lockless lookup replaces locking with sequence counter checks:
6358  *
6359  * vfs_smr_enter();
6360  * dvp_seqc = seqc_read_any(dvp);
6361  * // fail if someone is altering the directory vnode
6362  * if (seqc_in_modify(dvp_seqc))
6363  *     abort();
6364  * for (;;) {
6365  *      // permission check. note it can race, but we will validate the outcome
6366  *      // with a seqc
6367  * 	if (!canlookup_smr(dvp, cnp)) {
6368  * 	    // has dvp changed from under us? if so, the denial may be invalid
6369  *	    if (!seqc_consistent(dvp, dvp_seqc)
6370  * 	        fallback_to_locked();
6371  * 	    // nothing changed, lookup denial is valid
6372  * 	    fail();
6373  * 	}
6374  * 	// look for the target name inside dvp
6375  * 	tvp = findnext(dvp, cnp);
6376  * 	tvp_seqc = seqc_read_any(tvp);
6377  *	// bail if someone is altering the target vnode
6378  * 	if (seqc_in_modify(tvp_seqc))
6379  * 	    fallback_to_locked();
6380  *	// bail if someone is altering the directory vnode
6381  * 	if (!seqc_consistent(dvp, dvp_seqc)
6382  * 	    fallback_to_locked();
6383  * 	// we confirmed neither dvp nor tvp changed while we were making the
6384  * 	// jump to the next component, thus the result is the same as if we
6385  *      // held the lock on dvp and tvp the entire time, continue with tvp
6386  *      // as the directory to look up names in
6387  * 	dvp = tvp;
6388  * 	dvp_seqc = tvp_seqc;
6389  * 	if (last)
6390  * 	    break;
6391  * }
6392  * vget(); // secure the vnode
6393  * if (!seqc_consistent(tvp, tvp_seqc) // final check
6394  *     fallback_to_locked();
6395  * // at this point we know nothing has changed for any parent<->child pair
6396  * // as they were crossed during the lookup, meaning we matched the guarantee
6397  * // of the locked variant
6398  * return (tvp);
6399  *
6400  * The API contract for VOP_FPLOOKUP_VEXEC routines is as follows:
6401  * - they are called while within vfs_smr protection which they must never exit
6402  * - EAGAIN can be returned to denote checking could not be performed, it is
6403  *   always valid to return it
6404  * - if the sequence counter has not changed the result must be valid
6405  * - if the sequence counter has changed both false positives and false negatives
6406  *   are permitted (since the result will be rejected later)
6407  * - for simple cases of unix permission checks vaccess_vexec_smr can be used
6408  *
6409  * Caveats to watch out for:
6410  * - vnodes are passed unlocked and unreferenced with nothing stopping
6411  *   VOP_RECLAIM, in turn meaning that ->v_data can become NULL. It is advised
6412  *   to use atomic_load_ptr to fetch it.
6413  * - the aforementioned object can also get freed, meaning absent other means it
6414  *   should be protected with vfs_smr
6415  * - either safely checking permissions as they are modified or guaranteeing
6416  *   their stability is left to the routine
6417  */
6418 int
cache_fplookup(struct nameidata * ndp,enum cache_fpl_status * status,struct pwd ** pwdp)6419 cache_fplookup(struct nameidata *ndp, enum cache_fpl_status *status,
6420     struct pwd **pwdp)
6421 {
6422 	struct cache_fpl fpl;
6423 	struct pwd *pwd;
6424 	struct vnode *dvp;
6425 	struct componentname *cnp;
6426 	int error;
6427 
6428 	fpl.status = CACHE_FPL_STATUS_UNSET;
6429 	fpl.in_smr = false;
6430 	fpl.ndp = ndp;
6431 	fpl.cnp = cnp = &ndp->ni_cnd;
6432 	MPASS(ndp->ni_lcf == 0);
6433 	KASSERT ((cnp->cn_flags & CACHE_FPL_INTERNAL_CN_FLAGS) == 0,
6434 	    ("%s: internal flags found in cn_flags %" PRIx64, __func__,
6435 	    cnp->cn_flags));
6436 	MPASS(cnp->cn_nameptr == cnp->cn_pnbuf);
6437 	MPASS(ndp->ni_resflags == 0);
6438 
6439 	if (__predict_false(!cache_can_fplookup(&fpl))) {
6440 		*status = fpl.status;
6441 		SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
6442 		return (EOPNOTSUPP);
6443 	}
6444 
6445 	cache_fpl_checkpoint_outer(&fpl);
6446 
6447 	cache_fpl_smr_enter_initial(&fpl);
6448 #ifdef INVARIANTS
6449 	fpl.debug.ni_pathlen = ndp->ni_pathlen;
6450 #endif
6451 	fpl.nulchar = &cnp->cn_nameptr[ndp->ni_pathlen - 1];
6452 	fpl.fsearch = false;
6453 	fpl.tvp = NULL; /* for degenerate path handling */
6454 	fpl.pwd = pwdp;
6455 	pwd = pwd_get_smr();
6456 	*(fpl.pwd) = pwd;
6457 	namei_setup_rootdir(ndp, cnp, pwd);
6458 	ndp->ni_topdir = pwd->pwd_jdir;
6459 
6460 	if (cnp->cn_pnbuf[0] == '/') {
6461 		dvp = cache_fpl_handle_root(&fpl);
6462 		ndp->ni_resflags = NIRES_ABS;
6463 	} else {
6464 		if (ndp->ni_dirfd == AT_FDCWD) {
6465 			dvp = pwd->pwd_cdir;
6466 		} else {
6467 			error = cache_fplookup_dirfd(&fpl, &dvp);
6468 			if (__predict_false(error != 0)) {
6469 				goto out;
6470 			}
6471 		}
6472 	}
6473 
6474 	SDT_PROBE4(vfs, namei, lookup, entry, dvp, cnp->cn_pnbuf, cnp->cn_flags, true);
6475 	error = cache_fplookup_impl(dvp, &fpl);
6476 out:
6477 	cache_fpl_smr_assert_not_entered(&fpl);
6478 	cache_fpl_assert_status(&fpl);
6479 	*status = fpl.status;
6480 	SDT_PROBE3(vfs, fplookup, lookup, done, ndp, fpl.line, fpl.status);
6481 	if (__predict_true(fpl.status == CACHE_FPL_STATUS_HANDLED)) {
6482 		MPASS(error != CACHE_FPL_FAILED);
6483 		SDT_PROBE4(vfs, namei, lookup, return, error, ndp->ni_vp, true,
6484 		    ndp);
6485 		if (error != 0) {
6486 			cache_fpl_cleanup_cnp(fpl.cnp);
6487 			MPASS(fpl.dvp == NULL);
6488 			MPASS(fpl.tvp == NULL);
6489 		}
6490 		ndp->ni_dvp = fpl.dvp;
6491 		ndp->ni_vp = fpl.tvp;
6492 	}
6493 	return (error);
6494 }
6495