xref: /freebsd/sys/kern/vfs_subr.c (revision e60f608eb9cf3b38099948545934d699de9bbcea)
1 /*-
2  * SPDX-License-Identifier: BSD-3-Clause
3  *
4  * Copyright (c) 1989, 1993
5  *	The Regents of the University of California.  All rights reserved.
6  * (c) UNIX System Laboratories, Inc.
7  * All or some portions of this file are derived from material licensed
8  * to the University of California by American Telephone and Telegraph
9  * Co. or Unix System Laboratories, Inc. and are reproduced herein with
10  * the permission of UNIX System Laboratories, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  * 3. Neither the name of the University nor the names of its contributors
21  *    may be used to endorse or promote products derived from this software
22  *    without specific prior written permission.
23  *
24  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34  * SUCH DAMAGE.
35  */
36 
37 /*
38  * External virtual filesystem routines
39  */
40 
41 #include <sys/cdefs.h>
42 #include "opt_ddb.h"
43 #include "opt_watchdog.h"
44 
45 #include <sys/param.h>
46 #include <sys/systm.h>
47 #include <sys/asan.h>
48 #include <sys/bio.h>
49 #include <sys/buf.h>
50 #include <sys/capsicum.h>
51 #include <sys/condvar.h>
52 #include <sys/conf.h>
53 #include <sys/counter.h>
54 #include <sys/dirent.h>
55 #include <sys/event.h>
56 #include <sys/eventhandler.h>
57 #include <sys/extattr.h>
58 #include <sys/file.h>
59 #include <sys/fcntl.h>
60 #include <sys/jail.h>
61 #include <sys/kdb.h>
62 #include <sys/kernel.h>
63 #include <sys/kthread.h>
64 #include <sys/ktr.h>
65 #include <sys/limits.h>
66 #include <sys/lockf.h>
67 #include <sys/malloc.h>
68 #include <sys/mount.h>
69 #include <sys/namei.h>
70 #include <sys/pctrie.h>
71 #include <sys/priv.h>
72 #include <sys/reboot.h>
73 #include <sys/refcount.h>
74 #include <sys/rwlock.h>
75 #include <sys/sched.h>
76 #include <sys/sleepqueue.h>
77 #include <sys/smr.h>
78 #include <sys/smp.h>
79 #include <sys/stat.h>
80 #include <sys/sysctl.h>
81 #include <sys/syslog.h>
82 #include <sys/user.h>
83 #include <sys/vmmeter.h>
84 #include <sys/vnode.h>
85 #include <sys/watchdog.h>
86 
87 #include <machine/stdarg.h>
88 
89 #include <security/mac/mac_framework.h>
90 
91 #include <vm/vm.h>
92 #include <vm/vm_object.h>
93 #include <vm/vm_extern.h>
94 #include <vm/pmap.h>
95 #include <vm/vm_map.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_kern.h>
98 #include <vm/vnode_pager.h>
99 #include <vm/uma.h>
100 
101 #if defined(DEBUG_VFS_LOCKS) && (!defined(INVARIANTS) || !defined(WITNESS))
102 #error DEBUG_VFS_LOCKS requires INVARIANTS and WITNESS
103 #endif
104 
105 #ifdef DDB
106 #include <ddb/ddb.h>
107 #endif
108 
109 static void	delmntque(struct vnode *vp);
110 static int	flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
111 		    int slpflag, int slptimeo);
112 static void	syncer_shutdown(void *arg, int howto);
113 static int	vtryrecycle(struct vnode *vp, bool isvnlru);
114 static void	v_init_counters(struct vnode *);
115 static void	vn_seqc_init(struct vnode *);
116 static void	vn_seqc_write_end_free(struct vnode *vp);
117 static void	vgonel(struct vnode *);
118 static bool	vhold_recycle_free(struct vnode *);
119 static void	vdropl_recycle(struct vnode *vp);
120 static void	vdrop_recycle(struct vnode *vp);
121 static void	vfs_knllock(void *arg);
122 static void	vfs_knlunlock(void *arg);
123 static void	vfs_knl_assert_lock(void *arg, int what);
124 static void	destroy_vpollinfo(struct vpollinfo *vi);
125 static int	v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
126 		    daddr_t startlbn, daddr_t endlbn);
127 static void	vnlru_recalc(void);
128 
129 static SYSCTL_NODE(_vfs, OID_AUTO, vnode, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
130     "vnode configuration and statistics");
131 static SYSCTL_NODE(_vfs_vnode, OID_AUTO, param, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
132     "vnode configuration");
133 static SYSCTL_NODE(_vfs_vnode, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
134     "vnode statistics");
135 static SYSCTL_NODE(_vfs_vnode, OID_AUTO, vnlru, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
136     "vnode recycling");
137 
138 /*
139  * Number of vnodes in existence.  Increased whenever getnewvnode()
140  * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
141  */
142 static u_long __exclusive_cache_line numvnodes;
143 
144 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
145     "Number of vnodes in existence (legacy)");
146 SYSCTL_ULONG(_vfs_vnode_stats, OID_AUTO, count, CTLFLAG_RD, &numvnodes, 0,
147     "Number of vnodes in existence");
148 
149 static counter_u64_t vnodes_created;
150 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
151     "Number of vnodes created by getnewvnode (legacy)");
152 SYSCTL_COUNTER_U64(_vfs_vnode_stats, OID_AUTO, created, CTLFLAG_RD, &vnodes_created,
153     "Number of vnodes created by getnewvnode");
154 
155 /*
156  * Conversion tables for conversion from vnode types to inode formats
157  * and back.
158  */
159 __enum_uint8(vtype) iftovt_tab[16] = {
160 	VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
161 	VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
162 };
163 int vttoif_tab[10] = {
164 	0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
165 	S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
166 };
167 
168 /*
169  * List of allocates vnodes in the system.
170  */
171 static TAILQ_HEAD(freelst, vnode) vnode_list;
172 static struct vnode *vnode_list_free_marker;
173 static struct vnode *vnode_list_reclaim_marker;
174 
175 /*
176  * "Free" vnode target.  Free vnodes are rarely completely free, but are
177  * just ones that are cheap to recycle.  Usually they are for files which
178  * have been stat'd but not read; these usually have inode and namecache
179  * data attached to them.  This target is the preferred minimum size of a
180  * sub-cache consisting mostly of such files. The system balances the size
181  * of this sub-cache with its complement to try to prevent either from
182  * thrashing while the other is relatively inactive.  The targets express
183  * a preference for the best balance.
184  *
185  * "Above" this target there are 2 further targets (watermarks) related
186  * to recyling of free vnodes.  In the best-operating case, the cache is
187  * exactly full, the free list has size between vlowat and vhiwat above the
188  * free target, and recycling from it and normal use maintains this state.
189  * Sometimes the free list is below vlowat or even empty, but this state
190  * is even better for immediate use provided the cache is not full.
191  * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
192  * ones) to reach one of these states.  The watermarks are currently hard-
193  * coded as 4% and 9% of the available space higher.  These and the default
194  * of 25% for wantfreevnodes are too large if the memory size is large.
195  * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
196  * whenever vnlru_proc() becomes active.
197  */
198 static long wantfreevnodes;
199 static long __exclusive_cache_line freevnodes;
200 static long freevnodes_old;
201 
202 static u_long recycles_count;
203 SYSCTL_ULONG(_vfs, OID_AUTO, recycles, CTLFLAG_RD | CTLFLAG_STATS, &recycles_count, 0,
204     "Number of vnodes recycled to meet vnode cache targets (legacy)");
205 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, recycles, CTLFLAG_RD | CTLFLAG_STATS,
206     &recycles_count, 0,
207     "Number of vnodes recycled to meet vnode cache targets");
208 
209 static u_long recycles_free_count;
210 SYSCTL_ULONG(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD | CTLFLAG_STATS,
211     &recycles_free_count, 0,
212     "Number of free vnodes recycled to meet vnode cache targets (legacy)");
213 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, recycles_free, CTLFLAG_RD | CTLFLAG_STATS,
214     &recycles_free_count, 0,
215     "Number of free vnodes recycled to meet vnode cache targets");
216 
217 static counter_u64_t direct_recycles_free_count;
218 SYSCTL_COUNTER_U64(_vfs_vnode_vnlru, OID_AUTO, direct_recycles_free, CTLFLAG_RD,
219     &direct_recycles_free_count,
220     "Number of free vnodes recycled by vn_alloc callers to meet vnode cache targets");
221 
222 static counter_u64_t vnode_skipped_requeues;
223 SYSCTL_COUNTER_U64(_vfs_vnode_stats, OID_AUTO, skipped_requeues, CTLFLAG_RD, &vnode_skipped_requeues,
224     "Number of times LRU requeue was skipped due to lock contention");
225 
226 static __read_mostly bool vnode_can_skip_requeue;
227 SYSCTL_BOOL(_vfs_vnode_param, OID_AUTO, can_skip_requeue, CTLFLAG_RW,
228     &vnode_can_skip_requeue, 0, "Is LRU requeue skippable");
229 
230 static u_long deferred_inact;
231 SYSCTL_ULONG(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD,
232     &deferred_inact, 0, "Number of times inactive processing was deferred");
233 
234 /* To keep more than one thread at a time from running vfs_getnewfsid */
235 static struct mtx mntid_mtx;
236 
237 /*
238  * Lock for any access to the following:
239  *	vnode_list
240  *	numvnodes
241  *	freevnodes
242  */
243 static struct mtx __exclusive_cache_line vnode_list_mtx;
244 
245 /* Publicly exported FS */
246 struct nfs_public nfs_pub;
247 
248 static uma_zone_t buf_trie_zone;
249 static smr_t buf_trie_smr;
250 
251 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
252 static uma_zone_t vnode_zone;
253 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
254 
255 __read_frequently smr_t vfs_smr;
256 
257 /*
258  * The workitem queue.
259  *
260  * It is useful to delay writes of file data and filesystem metadata
261  * for tens of seconds so that quickly created and deleted files need
262  * not waste disk bandwidth being created and removed. To realize this,
263  * we append vnodes to a "workitem" queue. When running with a soft
264  * updates implementation, most pending metadata dependencies should
265  * not wait for more than a few seconds. Thus, mounted on block devices
266  * are delayed only about a half the time that file data is delayed.
267  * Similarly, directory updates are more critical, so are only delayed
268  * about a third the time that file data is delayed. Thus, there are
269  * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
270  * one each second (driven off the filesystem syncer process). The
271  * syncer_delayno variable indicates the next queue that is to be processed.
272  * Items that need to be processed soon are placed in this queue:
273  *
274  *	syncer_workitem_pending[syncer_delayno]
275  *
276  * A delay of fifteen seconds is done by placing the request fifteen
277  * entries later in the queue:
278  *
279  *	syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
280  *
281  */
282 static int syncer_delayno;
283 static long syncer_mask;
284 LIST_HEAD(synclist, bufobj);
285 static struct synclist *syncer_workitem_pending;
286 /*
287  * The sync_mtx protects:
288  *	bo->bo_synclist
289  *	sync_vnode_count
290  *	syncer_delayno
291  *	syncer_state
292  *	syncer_workitem_pending
293  *	syncer_worklist_len
294  *	rushjob
295  */
296 static struct mtx sync_mtx;
297 static struct cv sync_wakeup;
298 
299 #define SYNCER_MAXDELAY		32
300 static int syncer_maxdelay = SYNCER_MAXDELAY;	/* maximum delay time */
301 static int syncdelay = 30;		/* max time to delay syncing data */
302 static int filedelay = 30;		/* time to delay syncing files */
303 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
304     "Time to delay syncing files (in seconds)");
305 static int dirdelay = 29;		/* time to delay syncing directories */
306 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
307     "Time to delay syncing directories (in seconds)");
308 static int metadelay = 28;		/* time to delay syncing metadata */
309 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
310     "Time to delay syncing metadata (in seconds)");
311 static int rushjob;		/* number of slots to run ASAP */
312 static int stat_rush_requests;	/* number of times I/O speeded up */
313 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
314     "Number of times I/O speeded up (rush requests)");
315 
316 #define	VDBATCH_SIZE 8
317 struct vdbatch {
318 	u_int index;
319 	struct mtx lock;
320 	struct vnode *tab[VDBATCH_SIZE];
321 };
322 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
323 
324 static void	vdbatch_dequeue(struct vnode *vp);
325 
326 /*
327  * The syncer will require at least SYNCER_MAXDELAY iterations to shutdown;
328  * we probably don't want to pause for the whole second each time.
329  */
330 #define SYNCER_SHUTDOWN_SPEEDUP		32
331 static int sync_vnode_count;
332 static int syncer_worklist_len;
333 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
334     syncer_state;
335 
336 /* Target for maximum number of vnodes. */
337 u_long desiredvnodes;
338 static u_long gapvnodes;		/* gap between wanted and desired */
339 static u_long vhiwat;		/* enough extras after expansion */
340 static u_long vlowat;		/* minimal extras before expansion */
341 static bool vstir;		/* nonzero to stir non-free vnodes */
342 static volatile int vsmalltrigger = 8;	/* pref to keep if > this many pages */
343 
344 static u_long vnlru_read_freevnodes(void);
345 
346 /*
347  * Note that no attempt is made to sanitize these parameters.
348  */
349 static int
sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)350 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
351 {
352 	u_long val;
353 	int error;
354 
355 	val = desiredvnodes;
356 	error = sysctl_handle_long(oidp, &val, 0, req);
357 	if (error != 0 || req->newptr == NULL)
358 		return (error);
359 
360 	if (val == desiredvnodes)
361 		return (0);
362 	mtx_lock(&vnode_list_mtx);
363 	desiredvnodes = val;
364 	wantfreevnodes = desiredvnodes / 4;
365 	vnlru_recalc();
366 	mtx_unlock(&vnode_list_mtx);
367 	/*
368 	 * XXX There is no protection against multiple threads changing
369 	 * desiredvnodes at the same time. Locking above only helps vnlru and
370 	 * getnewvnode.
371 	 */
372 	vfs_hash_changesize(desiredvnodes);
373 	cache_changesize(desiredvnodes);
374 	return (0);
375 }
376 
377 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
378     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
379     "LU", "Target for maximum number of vnodes (legacy)");
380 SYSCTL_PROC(_vfs_vnode_param, OID_AUTO, limit,
381     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
382     "LU", "Target for maximum number of vnodes");
383 
384 static int
sysctl_freevnodes(SYSCTL_HANDLER_ARGS)385 sysctl_freevnodes(SYSCTL_HANDLER_ARGS)
386 {
387 	u_long rfreevnodes;
388 
389 	rfreevnodes = vnlru_read_freevnodes();
390 	return (sysctl_handle_long(oidp, &rfreevnodes, 0, req));
391 }
392 
393 SYSCTL_PROC(_vfs, OID_AUTO, freevnodes,
394     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_freevnodes,
395     "LU", "Number of \"free\" vnodes (legacy)");
396 SYSCTL_PROC(_vfs_vnode_stats, OID_AUTO, free,
397     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RD, NULL, 0, sysctl_freevnodes,
398     "LU", "Number of \"free\" vnodes");
399 
400 static int
sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)401 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
402 {
403 	u_long val;
404 	int error;
405 
406 	val = wantfreevnodes;
407 	error = sysctl_handle_long(oidp, &val, 0, req);
408 	if (error != 0 || req->newptr == NULL)
409 		return (error);
410 
411 	if (val == wantfreevnodes)
412 		return (0);
413 	mtx_lock(&vnode_list_mtx);
414 	wantfreevnodes = val;
415 	vnlru_recalc();
416 	mtx_unlock(&vnode_list_mtx);
417 	return (0);
418 }
419 
420 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
421     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
422     "LU", "Target for minimum number of \"free\" vnodes (legacy)");
423 SYSCTL_PROC(_vfs_vnode_param, OID_AUTO, wantfree,
424     CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
425     "LU", "Target for minimum number of \"free\" vnodes");
426 
427 static int vnlru_nowhere;
428 SYSCTL_INT(_vfs_vnode_vnlru, OID_AUTO, failed_runs, CTLFLAG_RD | CTLFLAG_STATS,
429     &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
430 
431 static int
sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)432 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
433 {
434 	struct vnode *vp;
435 	struct nameidata nd;
436 	char *buf;
437 	unsigned long ndflags;
438 	int error;
439 
440 	if (req->newptr == NULL)
441 		return (EINVAL);
442 	if (req->newlen >= PATH_MAX)
443 		return (E2BIG);
444 
445 	buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
446 	error = SYSCTL_IN(req, buf, req->newlen);
447 	if (error != 0)
448 		goto out;
449 
450 	buf[req->newlen] = '\0';
451 
452 	ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1;
453 	NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf);
454 	if ((error = namei(&nd)) != 0)
455 		goto out;
456 	vp = nd.ni_vp;
457 
458 	if (VN_IS_DOOMED(vp)) {
459 		/*
460 		 * This vnode is being recycled.  Return != 0 to let the caller
461 		 * know that the sysctl had no effect.  Return EAGAIN because a
462 		 * subsequent call will likely succeed (since namei will create
463 		 * a new vnode if necessary)
464 		 */
465 		error = EAGAIN;
466 		goto putvnode;
467 	}
468 
469 	vgone(vp);
470 putvnode:
471 	vput(vp);
472 	NDFREE_PNBUF(&nd);
473 out:
474 	free(buf, M_TEMP);
475 	return (error);
476 }
477 
478 static int
sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)479 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
480 {
481 	struct thread *td = curthread;
482 	struct vnode *vp;
483 	struct file *fp;
484 	int error;
485 	int fd;
486 
487 	if (req->newptr == NULL)
488 		return (EBADF);
489 
490         error = sysctl_handle_int(oidp, &fd, 0, req);
491         if (error != 0)
492                 return (error);
493 	error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
494 	if (error != 0)
495 		return (error);
496 	vp = fp->f_vnode;
497 
498 	error = vn_lock(vp, LK_EXCLUSIVE);
499 	if (error != 0)
500 		goto drop;
501 
502 	vgone(vp);
503 	VOP_UNLOCK(vp);
504 drop:
505 	fdrop(fp, td);
506 	return (error);
507 }
508 
509 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
510     CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
511     sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
512 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
513     CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
514     sysctl_ftry_reclaim_vnode, "I",
515     "Try to reclaim a vnode by its file descriptor");
516 
517 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
518 #define vnsz2log 8
519 #ifndef DEBUG_LOCKS
520 _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log &&
521     sizeof(struct vnode) < 1UL << (vnsz2log + 1),
522     "vnsz2log needs to be updated");
523 #endif
524 
525 /*
526  * Support for the bufobj clean & dirty pctrie.
527  */
528 static void *
buf_trie_alloc(struct pctrie * ptree)529 buf_trie_alloc(struct pctrie *ptree)
530 {
531 	return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
532 }
533 
534 static void
buf_trie_free(struct pctrie * ptree,void * node)535 buf_trie_free(struct pctrie *ptree, void *node)
536 {
537 	uma_zfree_smr(buf_trie_zone, node);
538 }
539 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
540     buf_trie_smr);
541 
542 /*
543  * Lookup the next element greater than or equal to lblkno, accounting for the
544  * fact that, for pctries, negative values are greater than nonnegative ones.
545  */
546 static struct buf *
buf_lookup_ge(struct bufv * bv,daddr_t lblkno)547 buf_lookup_ge(struct bufv *bv, daddr_t lblkno)
548 {
549 	struct buf *bp;
550 
551 	bp = BUF_PCTRIE_LOOKUP_GE(&bv->bv_root, lblkno);
552 	if (bp == NULL && lblkno < 0)
553 		bp = BUF_PCTRIE_LOOKUP_GE(&bv->bv_root, 0);
554 	if (bp != NULL && bp->b_lblkno < lblkno)
555 		bp = NULL;
556 	return (bp);
557 }
558 
559 /*
560  * Insert bp, and find the next element smaller than bp, accounting for the fact
561  * that, for pctries, negative values are greater than nonnegative ones.
562  */
563 static int
buf_insert_lookup_le(struct bufv * bv,struct buf * bp,struct buf ** n)564 buf_insert_lookup_le(struct bufv *bv, struct buf *bp, struct buf **n)
565 {
566 	int error;
567 
568 	error = BUF_PCTRIE_INSERT_LOOKUP_LE(&bv->bv_root, bp, n);
569 	if (error != EEXIST) {
570 		if (*n == NULL && bp->b_lblkno >= 0)
571 			*n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, ~0L);
572 		if (*n != NULL && (*n)->b_lblkno >= bp->b_lblkno)
573 			*n = NULL;
574 	}
575 	return (error);
576 }
577 
578 /*
579  * Initialize the vnode management data structures.
580  *
581  * Reevaluate the following cap on the number of vnodes after the physical
582  * memory size exceeds 512GB.  In the limit, as the physical memory size
583  * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
584  */
585 #ifndef	MAXVNODES_MAX
586 #define	MAXVNODES_MAX	(512UL * 1024 * 1024 / 64)	/* 8M */
587 #endif
588 
589 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
590 
591 static struct vnode *
vn_alloc_marker(struct mount * mp)592 vn_alloc_marker(struct mount *mp)
593 {
594 	struct vnode *vp;
595 
596 	vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
597 	vp->v_type = VMARKER;
598 	vp->v_mount = mp;
599 
600 	return (vp);
601 }
602 
603 static void
vn_free_marker(struct vnode * vp)604 vn_free_marker(struct vnode *vp)
605 {
606 
607 	MPASS(vp->v_type == VMARKER);
608 	free(vp, M_VNODE_MARKER);
609 }
610 
611 #ifdef KASAN
612 static int
vnode_ctor(void * mem,int size,void * arg __unused,int flags __unused)613 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
614 {
615 	kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
616 	return (0);
617 }
618 
619 static void
vnode_dtor(void * mem,int size,void * arg __unused)620 vnode_dtor(void *mem, int size, void *arg __unused)
621 {
622 	size_t end1, end2, off1, off2;
623 
624 	_Static_assert(offsetof(struct vnode, v_vnodelist) <
625 	    offsetof(struct vnode, v_dbatchcpu),
626 	    "KASAN marks require updating");
627 
628 	off1 = offsetof(struct vnode, v_vnodelist);
629 	off2 = offsetof(struct vnode, v_dbatchcpu);
630 	end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
631 	end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
632 
633 	/*
634 	 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
635 	 * after the vnode has been freed.  Try to get some KASAN coverage by
636 	 * marking everything except those two fields as invalid.  Because
637 	 * KASAN's tracking is not byte-granular, any preceding fields sharing
638 	 * the same 8-byte aligned word must also be marked valid.
639 	 */
640 
641 	/* Handle the area from the start until v_vnodelist... */
642 	off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
643 	kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
644 
645 	/* ... then the area between v_vnodelist and v_dbatchcpu ... */
646 	off1 = roundup2(end1, KASAN_SHADOW_SCALE);
647 	off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
648 	if (off2 > off1)
649 		kasan_mark((void *)((char *)mem + off1), off2 - off1,
650 		    off2 - off1, KASAN_UMA_FREED);
651 
652 	/* ... and finally the area from v_dbatchcpu to the end. */
653 	off2 = roundup2(end2, KASAN_SHADOW_SCALE);
654 	kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
655 	    KASAN_UMA_FREED);
656 }
657 #endif /* KASAN */
658 
659 /*
660  * Initialize a vnode as it first enters the zone.
661  */
662 static int
vnode_init(void * mem,int size,int flags)663 vnode_init(void *mem, int size, int flags)
664 {
665 	struct vnode *vp;
666 
667 	vp = mem;
668 	bzero(vp, size);
669 	/*
670 	 * Setup locks.
671 	 */
672 	vp->v_vnlock = &vp->v_lock;
673 	mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
674 	/*
675 	 * By default, don't allow shared locks unless filesystems opt-in.
676 	 */
677 	lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
678 	    LK_NOSHARE | LK_IS_VNODE);
679 	/*
680 	 * Initialize bufobj.
681 	 */
682 	bufobj_init(&vp->v_bufobj, vp);
683 	/*
684 	 * Initialize namecache.
685 	 */
686 	cache_vnode_init(vp);
687 	/*
688 	 * Initialize rangelocks.
689 	 */
690 	rangelock_init(&vp->v_rl);
691 
692 	vp->v_dbatchcpu = NOCPU;
693 
694 	vp->v_state = VSTATE_DEAD;
695 
696 	/*
697 	 * Check vhold_recycle_free for an explanation.
698 	 */
699 	vp->v_holdcnt = VHOLD_NO_SMR;
700 	vp->v_type = VNON;
701 	mtx_lock(&vnode_list_mtx);
702 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
703 	mtx_unlock(&vnode_list_mtx);
704 	return (0);
705 }
706 
707 /*
708  * Free a vnode when it is cleared from the zone.
709  */
710 static void
vnode_fini(void * mem,int size)711 vnode_fini(void *mem, int size)
712 {
713 	struct vnode *vp;
714 	struct bufobj *bo;
715 
716 	vp = mem;
717 	vdbatch_dequeue(vp);
718 	mtx_lock(&vnode_list_mtx);
719 	TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
720 	mtx_unlock(&vnode_list_mtx);
721 	rangelock_destroy(&vp->v_rl);
722 	lockdestroy(vp->v_vnlock);
723 	mtx_destroy(&vp->v_interlock);
724 	bo = &vp->v_bufobj;
725 	rw_destroy(BO_LOCKPTR(bo));
726 
727 	kasan_mark(mem, size, size, 0);
728 }
729 
730 /*
731  * Provide the size of NFS nclnode and NFS fh for calculation of the
732  * vnode memory consumption.  The size is specified directly to
733  * eliminate dependency on NFS-private header.
734  *
735  * Other filesystems may use bigger or smaller (like UFS and ZFS)
736  * private inode data, but the NFS-based estimation is ample enough.
737  * Still, we care about differences in the size between 64- and 32-bit
738  * platforms.
739  *
740  * Namecache structure size is heuristically
741  * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
742  */
743 #ifdef _LP64
744 #define	NFS_NCLNODE_SZ	(528 + 64)
745 #define	NC_SZ		148
746 #else
747 #define	NFS_NCLNODE_SZ	(360 + 32)
748 #define	NC_SZ		92
749 #endif
750 
751 static void
vntblinit(void * dummy __unused)752 vntblinit(void *dummy __unused)
753 {
754 	struct vdbatch *vd;
755 	uma_ctor ctor;
756 	uma_dtor dtor;
757 	int cpu, physvnodes, virtvnodes;
758 
759 	/*
760 	 * Desiredvnodes is a function of the physical memory size and the
761 	 * kernel's heap size.  Generally speaking, it scales with the
762 	 * physical memory size.  The ratio of desiredvnodes to the physical
763 	 * memory size is 1:16 until desiredvnodes exceeds 98,304.
764 	 * Thereafter, the
765 	 * marginal ratio of desiredvnodes to the physical memory size is
766 	 * 1:64.  However, desiredvnodes is limited by the kernel's heap
767 	 * size.  The memory required by desiredvnodes vnodes and vm objects
768 	 * must not exceed 1/10th of the kernel's heap size.
769 	 */
770 	physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
771 	    3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
772 	virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
773 	    sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
774 	desiredvnodes = min(physvnodes, virtvnodes);
775 	if (desiredvnodes > MAXVNODES_MAX) {
776 		if (bootverbose)
777 			printf("Reducing kern.maxvnodes %lu -> %lu\n",
778 			    desiredvnodes, MAXVNODES_MAX);
779 		desiredvnodes = MAXVNODES_MAX;
780 	}
781 	wantfreevnodes = desiredvnodes / 4;
782 	mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
783 	TAILQ_INIT(&vnode_list);
784 	mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
785 	/*
786 	 * The lock is taken to appease WITNESS.
787 	 */
788 	mtx_lock(&vnode_list_mtx);
789 	vnlru_recalc();
790 	mtx_unlock(&vnode_list_mtx);
791 	vnode_list_free_marker = vn_alloc_marker(NULL);
792 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
793 	vnode_list_reclaim_marker = vn_alloc_marker(NULL);
794 	TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
795 
796 #ifdef KASAN
797 	ctor = vnode_ctor;
798 	dtor = vnode_dtor;
799 #else
800 	ctor = NULL;
801 	dtor = NULL;
802 #endif
803 	vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
804 	    vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
805 	uma_zone_set_smr(vnode_zone, vfs_smr);
806 
807 	/*
808 	 * Preallocate enough nodes to support one-per buf so that
809 	 * we can not fail an insert.  reassignbuf() callers can not
810 	 * tolerate the insertion failure.
811 	 */
812 	buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
813 	    NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
814 	    UMA_ZONE_NOFREE | UMA_ZONE_SMR);
815 	buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
816 	uma_prealloc(buf_trie_zone, nbuf);
817 
818 	vnodes_created = counter_u64_alloc(M_WAITOK);
819 	direct_recycles_free_count = counter_u64_alloc(M_WAITOK);
820 	vnode_skipped_requeues = counter_u64_alloc(M_WAITOK);
821 
822 	/*
823 	 * Initialize the filesystem syncer.
824 	 */
825 	syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
826 	    &syncer_mask);
827 	syncer_maxdelay = syncer_mask + 1;
828 	mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
829 	cv_init(&sync_wakeup, "syncer");
830 
831 	CPU_FOREACH(cpu) {
832 		vd = DPCPU_ID_PTR((cpu), vd);
833 		bzero(vd, sizeof(*vd));
834 		mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
835 	}
836 }
837 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
838 
839 /*
840  * Mark a mount point as busy. Used to synchronize access and to delay
841  * unmounting. Eventually, mountlist_mtx is not released on failure.
842  *
843  * vfs_busy() is a custom lock, it can block the caller.
844  * vfs_busy() only sleeps if the unmount is active on the mount point.
845  * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
846  * vnode belonging to mp.
847  *
848  * Lookup uses vfs_busy() to traverse mount points.
849  * root fs			var fs
850  * / vnode lock		A	/ vnode lock (/var)		D
851  * /var vnode lock	B	/log vnode lock(/var/log)	E
852  * vfs_busy lock	C	vfs_busy lock			F
853  *
854  * Within each file system, the lock order is C->A->B and F->D->E.
855  *
856  * When traversing across mounts, the system follows that lock order:
857  *
858  *        C->A->B
859  *              |
860  *              +->F->D->E
861  *
862  * The lookup() process for namei("/var") illustrates the process:
863  *  1. VOP_LOOKUP() obtains B while A is held
864  *  2. vfs_busy() obtains a shared lock on F while A and B are held
865  *  3. vput() releases lock on B
866  *  4. vput() releases lock on A
867  *  5. VFS_ROOT() obtains lock on D while shared lock on F is held
868  *  6. vfs_unbusy() releases shared lock on F
869  *  7. vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
870  *     Attempt to lock A (instead of vp_crossmp) while D is held would
871  *     violate the global order, causing deadlocks.
872  *
873  * dounmount() locks B while F is drained.  Note that for stacked
874  * filesystems, D and B in the example above may be the same lock,
875  * which introdues potential lock order reversal deadlock between
876  * dounmount() and step 5 above.  These filesystems may avoid the LOR
877  * by setting VV_CROSSLOCK on the covered vnode so that lock B will
878  * remain held until after step 5.
879  */
880 int
vfs_busy(struct mount * mp,int flags)881 vfs_busy(struct mount *mp, int flags)
882 {
883 	struct mount_pcpu *mpcpu;
884 
885 	MPASS((flags & ~MBF_MASK) == 0);
886 	CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
887 
888 	if (vfs_op_thread_enter(mp, mpcpu)) {
889 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
890 		MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
891 		MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
892 		vfs_mp_count_add_pcpu(mpcpu, ref, 1);
893 		vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
894 		vfs_op_thread_exit(mp, mpcpu);
895 		if (flags & MBF_MNTLSTLOCK)
896 			mtx_unlock(&mountlist_mtx);
897 		return (0);
898 	}
899 
900 	MNT_ILOCK(mp);
901 	vfs_assert_mount_counters(mp);
902 	MNT_REF(mp);
903 	/*
904 	 * If mount point is currently being unmounted, sleep until the
905 	 * mount point fate is decided.  If thread doing the unmounting fails,
906 	 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
907 	 * that this mount point has survived the unmount attempt and vfs_busy
908 	 * should retry.  Otherwise the unmounter thread will set MNTK_REFEXPIRE
909 	 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
910 	 * about to be really destroyed.  vfs_busy needs to release its
911 	 * reference on the mount point in this case and return with ENOENT,
912 	 * telling the caller the mount it tried to busy is no longer valid.
913 	 */
914 	while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
915 		KASSERT(TAILQ_EMPTY(&mp->mnt_uppers),
916 		    ("%s: non-empty upper mount list with pending unmount",
917 		    __func__));
918 		if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
919 			MNT_REL(mp);
920 			MNT_IUNLOCK(mp);
921 			CTR1(KTR_VFS, "%s: failed busying before sleeping",
922 			    __func__);
923 			return (ENOENT);
924 		}
925 		if (flags & MBF_MNTLSTLOCK)
926 			mtx_unlock(&mountlist_mtx);
927 		mp->mnt_kern_flag |= MNTK_MWAIT;
928 		msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
929 		if (flags & MBF_MNTLSTLOCK)
930 			mtx_lock(&mountlist_mtx);
931 		MNT_ILOCK(mp);
932 	}
933 	if (flags & MBF_MNTLSTLOCK)
934 		mtx_unlock(&mountlist_mtx);
935 	mp->mnt_lockref++;
936 	MNT_IUNLOCK(mp);
937 	return (0);
938 }
939 
940 /*
941  * Free a busy filesystem.
942  */
943 void
vfs_unbusy(struct mount * mp)944 vfs_unbusy(struct mount *mp)
945 {
946 	struct mount_pcpu *mpcpu;
947 	int c;
948 
949 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
950 
951 	if (vfs_op_thread_enter(mp, mpcpu)) {
952 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
953 		vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
954 		vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
955 		vfs_op_thread_exit(mp, mpcpu);
956 		return;
957 	}
958 
959 	MNT_ILOCK(mp);
960 	vfs_assert_mount_counters(mp);
961 	MNT_REL(mp);
962 	c = --mp->mnt_lockref;
963 	if (mp->mnt_vfs_ops == 0) {
964 		MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
965 		MNT_IUNLOCK(mp);
966 		return;
967 	}
968 	if (c < 0)
969 		vfs_dump_mount_counters(mp);
970 	if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
971 		MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
972 		CTR1(KTR_VFS, "%s: waking up waiters", __func__);
973 		mp->mnt_kern_flag &= ~MNTK_DRAINING;
974 		wakeup(&mp->mnt_lockref);
975 	}
976 	MNT_IUNLOCK(mp);
977 }
978 
979 /*
980  * Lookup a mount point by filesystem identifier.
981  */
982 struct mount *
vfs_getvfs(fsid_t * fsid)983 vfs_getvfs(fsid_t *fsid)
984 {
985 	struct mount *mp;
986 
987 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
988 	mtx_lock(&mountlist_mtx);
989 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
990 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
991 			vfs_ref(mp);
992 			mtx_unlock(&mountlist_mtx);
993 			return (mp);
994 		}
995 	}
996 	mtx_unlock(&mountlist_mtx);
997 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
998 	return ((struct mount *) 0);
999 }
1000 
1001 /*
1002  * Lookup a mount point by filesystem identifier, busying it before
1003  * returning.
1004  *
1005  * To avoid congestion on mountlist_mtx, implement simple direct-mapped
1006  * cache for popular filesystem identifiers.  The cache is lockess, using
1007  * the fact that struct mount's are never freed.  In worst case we may
1008  * get pointer to unmounted or even different filesystem, so we have to
1009  * check what we got, and go slow way if so.
1010  */
1011 struct mount *
vfs_busyfs(fsid_t * fsid)1012 vfs_busyfs(fsid_t *fsid)
1013 {
1014 #define	FSID_CACHE_SIZE	256
1015 	typedef struct mount * volatile vmp_t;
1016 	static vmp_t cache[FSID_CACHE_SIZE];
1017 	struct mount *mp;
1018 	int error;
1019 	uint32_t hash;
1020 
1021 	CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
1022 	hash = fsid->val[0] ^ fsid->val[1];
1023 	hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
1024 	mp = cache[hash];
1025 	if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
1026 		goto slow;
1027 	if (vfs_busy(mp, 0) != 0) {
1028 		cache[hash] = NULL;
1029 		goto slow;
1030 	}
1031 	if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
1032 		return (mp);
1033 	else
1034 	    vfs_unbusy(mp);
1035 
1036 slow:
1037 	mtx_lock(&mountlist_mtx);
1038 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
1039 		if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
1040 			error = vfs_busy(mp, MBF_MNTLSTLOCK);
1041 			if (error) {
1042 				cache[hash] = NULL;
1043 				mtx_unlock(&mountlist_mtx);
1044 				return (NULL);
1045 			}
1046 			cache[hash] = mp;
1047 			return (mp);
1048 		}
1049 	}
1050 	CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
1051 	mtx_unlock(&mountlist_mtx);
1052 	return ((struct mount *) 0);
1053 }
1054 
1055 /*
1056  * Check if a user can access privileged mount options.
1057  */
1058 int
vfs_suser(struct mount * mp,struct thread * td)1059 vfs_suser(struct mount *mp, struct thread *td)
1060 {
1061 	int error;
1062 
1063 	if (jailed(td->td_ucred)) {
1064 		/*
1065 		 * If the jail of the calling thread lacks permission for
1066 		 * this type of file system, deny immediately.
1067 		 */
1068 		if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
1069 			return (EPERM);
1070 
1071 		/*
1072 		 * If the file system was mounted outside the jail of the
1073 		 * calling thread, deny immediately.
1074 		 */
1075 		if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
1076 			return (EPERM);
1077 	}
1078 
1079 	/*
1080 	 * If file system supports delegated administration, we don't check
1081 	 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
1082 	 * by the file system itself.
1083 	 * If this is not the user that did original mount, we check for
1084 	 * the PRIV_VFS_MOUNT_OWNER privilege.
1085 	 */
1086 	if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
1087 	    mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
1088 		if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
1089 			return (error);
1090 	}
1091 	return (0);
1092 }
1093 
1094 /*
1095  * Get a new unique fsid.  Try to make its val[0] unique, since this value
1096  * will be used to create fake device numbers for stat().  Also try (but
1097  * not so hard) make its val[0] unique mod 2^16, since some emulators only
1098  * support 16-bit device numbers.  We end up with unique val[0]'s for the
1099  * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1100  *
1101  * Keep in mind that several mounts may be running in parallel.  Starting
1102  * the search one past where the previous search terminated is both a
1103  * micro-optimization and a defense against returning the same fsid to
1104  * different mounts.
1105  */
1106 void
vfs_getnewfsid(struct mount * mp)1107 vfs_getnewfsid(struct mount *mp)
1108 {
1109 	static uint16_t mntid_base;
1110 	struct mount *nmp;
1111 	fsid_t tfsid;
1112 	int mtype;
1113 
1114 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1115 	mtx_lock(&mntid_mtx);
1116 	mtype = mp->mnt_vfc->vfc_typenum;
1117 	tfsid.val[1] = mtype;
1118 	mtype = (mtype & 0xFF) << 24;
1119 	for (;;) {
1120 		tfsid.val[0] = makedev(255,
1121 		    mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1122 		mntid_base++;
1123 		if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1124 			break;
1125 		vfs_rel(nmp);
1126 	}
1127 	mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1128 	mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1129 	mtx_unlock(&mntid_mtx);
1130 }
1131 
1132 /*
1133  * Knob to control the precision of file timestamps:
1134  *
1135  *   0 = seconds only; nanoseconds zeroed.
1136  *   1 = seconds and nanoseconds, accurate within 1/HZ.
1137  *   2 = seconds and nanoseconds, truncated to microseconds.
1138  * >=3 = seconds and nanoseconds, maximum precision.
1139  */
1140 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1141 
1142 static int timestamp_precision = TSP_USEC;
1143 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1144     &timestamp_precision, 0, "File timestamp precision (0: seconds, "
1145     "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1146     "3+: sec + ns (max. precision))");
1147 
1148 /*
1149  * Get a current timestamp.
1150  */
1151 void
vfs_timestamp(struct timespec * tsp)1152 vfs_timestamp(struct timespec *tsp)
1153 {
1154 	struct timeval tv;
1155 
1156 	switch (timestamp_precision) {
1157 	case TSP_SEC:
1158 		tsp->tv_sec = time_second;
1159 		tsp->tv_nsec = 0;
1160 		break;
1161 	case TSP_HZ:
1162 		getnanotime(tsp);
1163 		break;
1164 	case TSP_USEC:
1165 		microtime(&tv);
1166 		TIMEVAL_TO_TIMESPEC(&tv, tsp);
1167 		break;
1168 	case TSP_NSEC:
1169 	default:
1170 		nanotime(tsp);
1171 		break;
1172 	}
1173 }
1174 
1175 /*
1176  * Set vnode attributes to VNOVAL
1177  */
1178 void
vattr_null(struct vattr * vap)1179 vattr_null(struct vattr *vap)
1180 {
1181 
1182 	vap->va_type = VNON;
1183 	vap->va_size = VNOVAL;
1184 	vap->va_bytes = VNOVAL;
1185 	vap->va_mode = VNOVAL;
1186 	vap->va_nlink = VNOVAL;
1187 	vap->va_uid = VNOVAL;
1188 	vap->va_gid = VNOVAL;
1189 	vap->va_fsid = VNOVAL;
1190 	vap->va_fileid = VNOVAL;
1191 	vap->va_blocksize = VNOVAL;
1192 	vap->va_rdev = VNOVAL;
1193 	vap->va_atime.tv_sec = VNOVAL;
1194 	vap->va_atime.tv_nsec = VNOVAL;
1195 	vap->va_mtime.tv_sec = VNOVAL;
1196 	vap->va_mtime.tv_nsec = VNOVAL;
1197 	vap->va_ctime.tv_sec = VNOVAL;
1198 	vap->va_ctime.tv_nsec = VNOVAL;
1199 	vap->va_birthtime.tv_sec = VNOVAL;
1200 	vap->va_birthtime.tv_nsec = VNOVAL;
1201 	vap->va_flags = VNOVAL;
1202 	vap->va_gen = VNOVAL;
1203 	vap->va_vaflags = 0;
1204 	vap->va_filerev = VNOVAL;
1205 }
1206 
1207 /*
1208  * Try to reduce the total number of vnodes.
1209  *
1210  * This routine (and its user) are buggy in at least the following ways:
1211  * - all parameters were picked years ago when RAM sizes were significantly
1212  *   smaller
1213  * - it can pick vnodes based on pages used by the vm object, but filesystems
1214  *   like ZFS don't use it making the pick broken
1215  * - since ZFS has its own aging policy it gets partially combated by this one
1216  * - a dedicated method should be provided for filesystems to let them decide
1217  *   whether the vnode should be recycled
1218  *
1219  * This routine is called when we have too many vnodes.  It attempts
1220  * to free <count> vnodes and will potentially free vnodes that still
1221  * have VM backing store (VM backing store is typically the cause
1222  * of a vnode blowout so we want to do this).  Therefore, this operation
1223  * is not considered cheap.
1224  *
1225  * A number of conditions may prevent a vnode from being reclaimed.
1226  * the buffer cache may have references on the vnode, a directory
1227  * vnode may still have references due to the namei cache representing
1228  * underlying files, or the vnode may be in active use.   It is not
1229  * desirable to reuse such vnodes.  These conditions may cause the
1230  * number of vnodes to reach some minimum value regardless of what
1231  * you set kern.maxvnodes to.  Do not set kern.maxvnodes too low.
1232  *
1233  * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1234  * 			 entries if this argument is strue
1235  * @param trigger	 Only reclaim vnodes with fewer than this many resident
1236  *			 pages.
1237  * @param target	 How many vnodes to reclaim.
1238  * @return		 The number of vnodes that were reclaimed.
1239  */
1240 static int
vlrureclaim(bool reclaim_nc_src,int trigger,u_long target)1241 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1242 {
1243 	struct vnode *vp, *mvp;
1244 	struct mount *mp;
1245 	struct vm_object *object;
1246 	u_long done;
1247 	bool retried;
1248 
1249 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1250 
1251 	retried = false;
1252 	done = 0;
1253 
1254 	mvp = vnode_list_reclaim_marker;
1255 restart:
1256 	vp = mvp;
1257 	while (done < target) {
1258 		vp = TAILQ_NEXT(vp, v_vnodelist);
1259 		if (__predict_false(vp == NULL))
1260 			break;
1261 
1262 		if (__predict_false(vp->v_type == VMARKER))
1263 			continue;
1264 
1265 		/*
1266 		 * If it's been deconstructed already, it's still
1267 		 * referenced, or it exceeds the trigger, skip it.
1268 		 * Also skip free vnodes.  We are trying to make space
1269 		 * for more free vnodes, not reduce their count.
1270 		 */
1271 		if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1272 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1273 			goto next_iter;
1274 
1275 		if (vp->v_type == VBAD || vp->v_type == VNON)
1276 			goto next_iter;
1277 
1278 		object = atomic_load_ptr(&vp->v_object);
1279 		if (object == NULL || object->resident_page_count > trigger) {
1280 			goto next_iter;
1281 		}
1282 
1283 		/*
1284 		 * Handle races against vnode allocation. Filesystems lock the
1285 		 * vnode some time after it gets returned from getnewvnode,
1286 		 * despite type and hold count being manipulated earlier.
1287 		 * Resorting to checking v_mount restores guarantees present
1288 		 * before the global list was reworked to contain all vnodes.
1289 		 */
1290 		if (!VI_TRYLOCK(vp))
1291 			goto next_iter;
1292 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1293 			VI_UNLOCK(vp);
1294 			goto next_iter;
1295 		}
1296 		if (vp->v_mount == NULL) {
1297 			VI_UNLOCK(vp);
1298 			goto next_iter;
1299 		}
1300 		vholdl(vp);
1301 		VI_UNLOCK(vp);
1302 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1303 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1304 		mtx_unlock(&vnode_list_mtx);
1305 
1306 		if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1307 			vdrop_recycle(vp);
1308 			goto next_iter_unlocked;
1309 		}
1310 		if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1311 			vdrop_recycle(vp);
1312 			vn_finished_write(mp);
1313 			goto next_iter_unlocked;
1314 		}
1315 
1316 		VI_LOCK(vp);
1317 		if (vp->v_usecount > 0 ||
1318 		    (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1319 		    (vp->v_object != NULL && vp->v_object->handle == vp &&
1320 		    vp->v_object->resident_page_count > trigger)) {
1321 			VOP_UNLOCK(vp);
1322 			vdropl_recycle(vp);
1323 			vn_finished_write(mp);
1324 			goto next_iter_unlocked;
1325 		}
1326 		recycles_count++;
1327 		vgonel(vp);
1328 		VOP_UNLOCK(vp);
1329 		vdropl_recycle(vp);
1330 		vn_finished_write(mp);
1331 		done++;
1332 next_iter_unlocked:
1333 		maybe_yield();
1334 		mtx_lock(&vnode_list_mtx);
1335 		goto restart;
1336 next_iter:
1337 		MPASS(vp->v_type != VMARKER);
1338 		if (!should_yield())
1339 			continue;
1340 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1341 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1342 		mtx_unlock(&vnode_list_mtx);
1343 		kern_yield(PRI_USER);
1344 		mtx_lock(&vnode_list_mtx);
1345 		goto restart;
1346 	}
1347 	if (done == 0 && !retried) {
1348 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1349 		TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1350 		retried = true;
1351 		goto restart;
1352 	}
1353 	return (done);
1354 }
1355 
1356 static int max_free_per_call = 10000;
1357 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_free_per_call, 0,
1358     "limit on vnode free requests per call to the vnlru_free routine (legacy)");
1359 SYSCTL_INT(_vfs_vnode_vnlru, OID_AUTO, max_free_per_call, CTLFLAG_RW,
1360     &max_free_per_call, 0,
1361     "limit on vnode free requests per call to the vnlru_free routine");
1362 
1363 /*
1364  * Attempt to recycle requested amount of free vnodes.
1365  */
1366 static int
vnlru_free_impl(int count,struct vfsops * mnt_op,struct vnode * mvp,bool isvnlru)1367 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp, bool isvnlru)
1368 {
1369 	struct vnode *vp;
1370 	struct mount *mp;
1371 	int ocount;
1372 	bool retried;
1373 
1374 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1375 	if (count > max_free_per_call)
1376 		count = max_free_per_call;
1377 	if (count == 0) {
1378 		mtx_unlock(&vnode_list_mtx);
1379 		return (0);
1380 	}
1381 	ocount = count;
1382 	retried = false;
1383 	vp = mvp;
1384 	for (;;) {
1385 		vp = TAILQ_NEXT(vp, v_vnodelist);
1386 		if (__predict_false(vp == NULL)) {
1387 			/*
1388 			 * The free vnode marker can be past eligible vnodes:
1389 			 * 1. if vdbatch_process trylock failed
1390 			 * 2. if vtryrecycle failed
1391 			 *
1392 			 * If so, start the scan from scratch.
1393 			 */
1394 			if (!retried && vnlru_read_freevnodes() > 0) {
1395 				TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1396 				TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1397 				vp = mvp;
1398 				retried = true;
1399 				continue;
1400 			}
1401 
1402 			/*
1403 			 * Give up
1404 			 */
1405 			TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1406 			TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1407 			mtx_unlock(&vnode_list_mtx);
1408 			break;
1409 		}
1410 		if (__predict_false(vp->v_type == VMARKER))
1411 			continue;
1412 		if (vp->v_holdcnt > 0)
1413 			continue;
1414 		/*
1415 		 * Don't recycle if our vnode is from different type
1416 		 * of mount point.  Note that mp is type-safe, the
1417 		 * check does not reach unmapped address even if
1418 		 * vnode is reclaimed.
1419 		 */
1420 		if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1421 		    mp->mnt_op != mnt_op) {
1422 			continue;
1423 		}
1424 		if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1425 			continue;
1426 		}
1427 		if (!vhold_recycle_free(vp))
1428 			continue;
1429 		TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1430 		TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1431 		mtx_unlock(&vnode_list_mtx);
1432 		/*
1433 		 * FIXME: ignores the return value, meaning it may be nothing
1434 		 * got recycled but it claims otherwise to the caller.
1435 		 *
1436 		 * Originally the value started being ignored in 2005 with
1437 		 * 114a1006a8204aa156e1f9ad6476cdff89cada7f .
1438 		 *
1439 		 * Respecting the value can run into significant stalls if most
1440 		 * vnodes belong to one file system and it has writes
1441 		 * suspended.  In presence of many threads and millions of
1442 		 * vnodes they keep contending on the vnode_list_mtx lock only
1443 		 * to find vnodes they can't recycle.
1444 		 *
1445 		 * The solution would be to pre-check if the vnode is likely to
1446 		 * be recycle-able, but it needs to happen with the
1447 		 * vnode_list_mtx lock held. This runs into a problem where
1448 		 * VOP_GETWRITEMOUNT (currently needed to find out about if
1449 		 * writes are frozen) can take locks which LOR against it.
1450 		 *
1451 		 * Check nullfs for one example (null_getwritemount).
1452 		 */
1453 		vtryrecycle(vp, isvnlru);
1454 		count--;
1455 		if (count == 0) {
1456 			break;
1457 		}
1458 		mtx_lock(&vnode_list_mtx);
1459 		vp = mvp;
1460 	}
1461 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1462 	return (ocount - count);
1463 }
1464 
1465 /*
1466  * XXX: returns without vnode_list_mtx locked!
1467  */
1468 static int
vnlru_free_locked_direct(int count)1469 vnlru_free_locked_direct(int count)
1470 {
1471 	int ret;
1472 
1473 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1474 	ret = vnlru_free_impl(count, NULL, vnode_list_free_marker, false);
1475 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1476 	return (ret);
1477 }
1478 
1479 static int
vnlru_free_locked_vnlru(int count)1480 vnlru_free_locked_vnlru(int count)
1481 {
1482 	int ret;
1483 
1484 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1485 	ret = vnlru_free_impl(count, NULL, vnode_list_free_marker, true);
1486 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1487 	return (ret);
1488 }
1489 
1490 static int
vnlru_free_vnlru(int count)1491 vnlru_free_vnlru(int count)
1492 {
1493 
1494 	mtx_lock(&vnode_list_mtx);
1495 	return (vnlru_free_locked_vnlru(count));
1496 }
1497 
1498 void
vnlru_free_vfsops(int count,struct vfsops * mnt_op,struct vnode * mvp)1499 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1500 {
1501 
1502 	MPASS(mnt_op != NULL);
1503 	MPASS(mvp != NULL);
1504 	VNPASS(mvp->v_type == VMARKER, mvp);
1505 	mtx_lock(&vnode_list_mtx);
1506 	vnlru_free_impl(count, mnt_op, mvp, true);
1507 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1508 }
1509 
1510 struct vnode *
vnlru_alloc_marker(void)1511 vnlru_alloc_marker(void)
1512 {
1513 	struct vnode *mvp;
1514 
1515 	mvp = vn_alloc_marker(NULL);
1516 	mtx_lock(&vnode_list_mtx);
1517 	TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1518 	mtx_unlock(&vnode_list_mtx);
1519 	return (mvp);
1520 }
1521 
1522 void
vnlru_free_marker(struct vnode * mvp)1523 vnlru_free_marker(struct vnode *mvp)
1524 {
1525 	mtx_lock(&vnode_list_mtx);
1526 	TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1527 	mtx_unlock(&vnode_list_mtx);
1528 	vn_free_marker(mvp);
1529 }
1530 
1531 static void
vnlru_recalc(void)1532 vnlru_recalc(void)
1533 {
1534 
1535 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1536 	gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1537 	vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1538 	vlowat = vhiwat / 2;
1539 }
1540 
1541 /*
1542  * Attempt to recycle vnodes in a context that is always safe to block.
1543  * Calling vlrurecycle() from the bowels of filesystem code has some
1544  * interesting deadlock problems.
1545  */
1546 static struct proc *vnlruproc;
1547 static int vnlruproc_sig;
1548 static u_long vnlruproc_kicks;
1549 
1550 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, kicks, CTLFLAG_RD, &vnlruproc_kicks, 0,
1551     "Number of times vnlru awakened due to vnode shortage");
1552 
1553 #define VNLRU_COUNT_SLOP 100
1554 
1555 /*
1556  * The main freevnodes counter is only updated when a counter local to CPU
1557  * diverges from 0 by more than VNLRU_FREEVNODES_SLOP. CPUs are conditionally
1558  * walked to compute a more accurate total.
1559  *
1560  * Note: the actual value at any given moment can still exceed slop, but it
1561  * should not be by significant margin in practice.
1562  */
1563 #define VNLRU_FREEVNODES_SLOP 126
1564 
1565 static void __noinline
vfs_freevnodes_rollup(int8_t * lfreevnodes)1566 vfs_freevnodes_rollup(int8_t *lfreevnodes)
1567 {
1568 
1569 	atomic_add_long(&freevnodes, *lfreevnodes);
1570 	*lfreevnodes = 0;
1571 	critical_exit();
1572 }
1573 
1574 static __inline void
vfs_freevnodes_inc(void)1575 vfs_freevnodes_inc(void)
1576 {
1577 	int8_t *lfreevnodes;
1578 
1579 	critical_enter();
1580 	lfreevnodes = PCPU_PTR(vfs_freevnodes);
1581 	(*lfreevnodes)++;
1582 	if (__predict_false(*lfreevnodes == VNLRU_FREEVNODES_SLOP))
1583 		vfs_freevnodes_rollup(lfreevnodes);
1584 	else
1585 		critical_exit();
1586 }
1587 
1588 static __inline void
vfs_freevnodes_dec(void)1589 vfs_freevnodes_dec(void)
1590 {
1591 	int8_t *lfreevnodes;
1592 
1593 	critical_enter();
1594 	lfreevnodes = PCPU_PTR(vfs_freevnodes);
1595 	(*lfreevnodes)--;
1596 	if (__predict_false(*lfreevnodes == -VNLRU_FREEVNODES_SLOP))
1597 		vfs_freevnodes_rollup(lfreevnodes);
1598 	else
1599 		critical_exit();
1600 }
1601 
1602 static u_long
vnlru_read_freevnodes(void)1603 vnlru_read_freevnodes(void)
1604 {
1605 	long slop, rfreevnodes, rfreevnodes_old;
1606 	int cpu;
1607 
1608 	rfreevnodes = atomic_load_long(&freevnodes);
1609 	rfreevnodes_old = atomic_load_long(&freevnodes_old);
1610 
1611 	if (rfreevnodes > rfreevnodes_old)
1612 		slop = rfreevnodes - rfreevnodes_old;
1613 	else
1614 		slop = rfreevnodes_old - rfreevnodes;
1615 	if (slop < VNLRU_FREEVNODES_SLOP)
1616 		return (rfreevnodes >= 0 ? rfreevnodes : 0);
1617 	CPU_FOREACH(cpu) {
1618 		rfreevnodes += cpuid_to_pcpu[cpu]->pc_vfs_freevnodes;
1619 	}
1620 	atomic_store_long(&freevnodes_old, rfreevnodes);
1621 	return (freevnodes_old >= 0 ? freevnodes_old : 0);
1622 }
1623 
1624 static bool
vnlru_under(u_long rnumvnodes,u_long limit)1625 vnlru_under(u_long rnumvnodes, u_long limit)
1626 {
1627 	u_long rfreevnodes, space;
1628 
1629 	if (__predict_false(rnumvnodes > desiredvnodes))
1630 		return (true);
1631 
1632 	space = desiredvnodes - rnumvnodes;
1633 	if (space < limit) {
1634 		rfreevnodes = vnlru_read_freevnodes();
1635 		if (rfreevnodes > wantfreevnodes)
1636 			space += rfreevnodes - wantfreevnodes;
1637 	}
1638 	return (space < limit);
1639 }
1640 
1641 static void
vnlru_kick_locked(void)1642 vnlru_kick_locked(void)
1643 {
1644 
1645 	mtx_assert(&vnode_list_mtx, MA_OWNED);
1646 	if (vnlruproc_sig == 0) {
1647 		vnlruproc_sig = 1;
1648 		vnlruproc_kicks++;
1649 		wakeup(vnlruproc);
1650 	}
1651 }
1652 
1653 static void
vnlru_kick_cond(void)1654 vnlru_kick_cond(void)
1655 {
1656 
1657 	if (vnlru_read_freevnodes() > wantfreevnodes)
1658 		return;
1659 
1660 	if (vnlruproc_sig)
1661 		return;
1662 	mtx_lock(&vnode_list_mtx);
1663 	vnlru_kick_locked();
1664 	mtx_unlock(&vnode_list_mtx);
1665 }
1666 
1667 static void
vnlru_proc_sleep(void)1668 vnlru_proc_sleep(void)
1669 {
1670 
1671 	if (vnlruproc_sig) {
1672 		vnlruproc_sig = 0;
1673 		wakeup(&vnlruproc_sig);
1674 	}
1675 	msleep(vnlruproc, &vnode_list_mtx, PVFS|PDROP, "vlruwt", hz);
1676 }
1677 
1678 /*
1679  * A lighter version of the machinery below.
1680  *
1681  * Tries to reach goals only by recycling free vnodes and does not invoke
1682  * uma_reclaim(UMA_RECLAIM_DRAIN).
1683  *
1684  * This works around pathological behavior in vnlru in presence of tons of free
1685  * vnodes, but without having to rewrite the machinery at this time. Said
1686  * behavior boils down to continuously trying to reclaim all kinds of vnodes
1687  * (cycling through all levels of "force") when the count is transiently above
1688  * limit. This happens a lot when all vnodes are used up and vn_alloc
1689  * speculatively increments the counter.
1690  *
1691  * Sample testcase: vnode limit 8388608, 20 separate directory trees each with
1692  * 1 million files in total and 20 find(1) processes stating them in parallel
1693  * (one per each tree).
1694  *
1695  * On a kernel with only stock machinery this needs anywhere between 60 and 120
1696  * seconds to execute (time varies *wildly* between runs). With the workaround
1697  * it consistently stays around 20 seconds [it got further down with later
1698  * changes].
1699  *
1700  * That is to say the entire thing needs a fundamental redesign (most notably
1701  * to accommodate faster recycling), the above only tries to get it ouf the way.
1702  *
1703  * Return values are:
1704  * -1 -- fallback to regular vnlru loop
1705  *  0 -- do nothing, go to sleep
1706  * >0 -- recycle this many vnodes
1707  */
1708 static long
vnlru_proc_light_pick(void)1709 vnlru_proc_light_pick(void)
1710 {
1711 	u_long rnumvnodes, rfreevnodes;
1712 
1713 	if (vstir || vnlruproc_sig == 1)
1714 		return (-1);
1715 
1716 	rnumvnodes = atomic_load_long(&numvnodes);
1717 	rfreevnodes = vnlru_read_freevnodes();
1718 
1719 	/*
1720 	 * vnode limit might have changed and now we may be at a significant
1721 	 * excess. Bail if we can't sort it out with free vnodes.
1722 	 *
1723 	 * Due to atomic updates the count can legitimately go above
1724 	 * the limit for a short period, don't bother doing anything in
1725 	 * that case.
1726 	 */
1727 	if (rnumvnodes > desiredvnodes + VNLRU_COUNT_SLOP + 10) {
1728 		if (rnumvnodes - rfreevnodes >= desiredvnodes ||
1729 		    rfreevnodes <= wantfreevnodes) {
1730 			return (-1);
1731 		}
1732 
1733 		return (rnumvnodes - desiredvnodes);
1734 	}
1735 
1736 	/*
1737 	 * Don't try to reach wantfreevnodes target if there are too few vnodes
1738 	 * to begin with.
1739 	 */
1740 	if (rnumvnodes < wantfreevnodes) {
1741 		return (0);
1742 	}
1743 
1744 	if (rfreevnodes < wantfreevnodes) {
1745 		return (-1);
1746 	}
1747 
1748 	return (0);
1749 }
1750 
1751 static bool
vnlru_proc_light(void)1752 vnlru_proc_light(void)
1753 {
1754 	long freecount;
1755 
1756 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1757 
1758 	freecount = vnlru_proc_light_pick();
1759 	if (freecount == -1)
1760 		return (false);
1761 
1762 	if (freecount != 0) {
1763 		vnlru_free_vnlru(freecount);
1764 	}
1765 
1766 	mtx_lock(&vnode_list_mtx);
1767 	vnlru_proc_sleep();
1768 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
1769 	return (true);
1770 }
1771 
1772 static u_long uma_reclaim_calls;
1773 SYSCTL_ULONG(_vfs_vnode_vnlru, OID_AUTO, uma_reclaim_calls, CTLFLAG_RD | CTLFLAG_STATS,
1774     &uma_reclaim_calls, 0, "Number of calls to uma_reclaim");
1775 
1776 static void
vnlru_proc(void)1777 vnlru_proc(void)
1778 {
1779 	u_long rnumvnodes, rfreevnodes, target;
1780 	unsigned long onumvnodes;
1781 	int done, force, trigger, usevnodes;
1782 	bool reclaim_nc_src, want_reread;
1783 
1784 	EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1785 	    SHUTDOWN_PRI_FIRST);
1786 
1787 	force = 0;
1788 	want_reread = false;
1789 	for (;;) {
1790 		kproc_suspend_check(vnlruproc);
1791 
1792 		if (force == 0 && vnlru_proc_light())
1793 			continue;
1794 
1795 		mtx_lock(&vnode_list_mtx);
1796 		rnumvnodes = atomic_load_long(&numvnodes);
1797 
1798 		if (want_reread) {
1799 			force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1800 			want_reread = false;
1801 		}
1802 
1803 		/*
1804 		 * If numvnodes is too large (due to desiredvnodes being
1805 		 * adjusted using its sysctl, or emergency growth), first
1806 		 * try to reduce it by discarding free vnodes.
1807 		 */
1808 		if (rnumvnodes > desiredvnodes + 10) {
1809 			vnlru_free_locked_vnlru(rnumvnodes - desiredvnodes);
1810 			mtx_lock(&vnode_list_mtx);
1811 			rnumvnodes = atomic_load_long(&numvnodes);
1812 		}
1813 		/*
1814 		 * Sleep if the vnode cache is in a good state.  This is
1815 		 * when it is not over-full and has space for about a 4%
1816 		 * or 9% expansion (by growing its size or inexcessively
1817 		 * reducing free vnode count).  Otherwise, try to reclaim
1818 		 * space for a 10% expansion.
1819 		 */
1820 		if (vstir && force == 0) {
1821 			force = 1;
1822 			vstir = false;
1823 		}
1824 		if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1825 			vnlru_proc_sleep();
1826 			continue;
1827 		}
1828 		rfreevnodes = vnlru_read_freevnodes();
1829 
1830 		onumvnodes = rnumvnodes;
1831 		/*
1832 		 * Calculate parameters for recycling.  These are the same
1833 		 * throughout the loop to give some semblance of fairness.
1834 		 * The trigger point is to avoid recycling vnodes with lots
1835 		 * of resident pages.  We aren't trying to free memory; we
1836 		 * are trying to recycle or at least free vnodes.
1837 		 */
1838 		if (rnumvnodes <= desiredvnodes)
1839 			usevnodes = rnumvnodes - rfreevnodes;
1840 		else
1841 			usevnodes = rnumvnodes;
1842 		if (usevnodes <= 0)
1843 			usevnodes = 1;
1844 		/*
1845 		 * The trigger value is chosen to give a conservatively
1846 		 * large value to ensure that it alone doesn't prevent
1847 		 * making progress.  The value can easily be so large that
1848 		 * it is effectively infinite in some congested and
1849 		 * misconfigured cases, and this is necessary.  Normally
1850 		 * it is about 8 to 100 (pages), which is quite large.
1851 		 */
1852 		trigger = vm_cnt.v_page_count * 2 / usevnodes;
1853 		if (force < 2)
1854 			trigger = vsmalltrigger;
1855 		reclaim_nc_src = force >= 3;
1856 		target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1857 		target = target / 10 + 1;
1858 		done = vlrureclaim(reclaim_nc_src, trigger, target);
1859 		mtx_unlock(&vnode_list_mtx);
1860 		/*
1861 		 * Total number of vnodes can transiently go slightly above the
1862 		 * limit (see vn_alloc_hard), no need to call uma_reclaim if
1863 		 * this happens.
1864 		 */
1865 		if (onumvnodes + VNLRU_COUNT_SLOP + 1000 > desiredvnodes &&
1866 		    numvnodes <= desiredvnodes) {
1867 			uma_reclaim_calls++;
1868 			uma_reclaim(UMA_RECLAIM_DRAIN);
1869 		}
1870 		if (done == 0) {
1871 			if (force == 0 || force == 1) {
1872 				force = 2;
1873 				continue;
1874 			}
1875 			if (force == 2) {
1876 				force = 3;
1877 				continue;
1878 			}
1879 			want_reread = true;
1880 			force = 0;
1881 			vnlru_nowhere++;
1882 			tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1883 		} else {
1884 			want_reread = true;
1885 			kern_yield(PRI_USER);
1886 		}
1887 	}
1888 }
1889 
1890 static struct kproc_desc vnlru_kp = {
1891 	"vnlru",
1892 	vnlru_proc,
1893 	&vnlruproc
1894 };
1895 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1896     &vnlru_kp);
1897 
1898 /*
1899  * Routines having to do with the management of the vnode table.
1900  */
1901 
1902 /*
1903  * Try to recycle a freed vnode.
1904  */
1905 static int
vtryrecycle(struct vnode * vp,bool isvnlru)1906 vtryrecycle(struct vnode *vp, bool isvnlru)
1907 {
1908 	struct mount *vnmp;
1909 
1910 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1911 	VNPASS(vp->v_holdcnt > 0, vp);
1912 	/*
1913 	 * This vnode may found and locked via some other list, if so we
1914 	 * can't recycle it yet.
1915 	 */
1916 	if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1917 		CTR2(KTR_VFS,
1918 		    "%s: impossible to recycle, vp %p lock is already held",
1919 		    __func__, vp);
1920 		vdrop_recycle(vp);
1921 		return (EWOULDBLOCK);
1922 	}
1923 	/*
1924 	 * Don't recycle if its filesystem is being suspended.
1925 	 */
1926 	if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1927 		VOP_UNLOCK(vp);
1928 		CTR2(KTR_VFS,
1929 		    "%s: impossible to recycle, cannot start the write for %p",
1930 		    __func__, vp);
1931 		vdrop_recycle(vp);
1932 		return (EBUSY);
1933 	}
1934 	/*
1935 	 * If we got this far, we need to acquire the interlock and see if
1936 	 * anyone picked up this vnode from another list.  If not, we will
1937 	 * mark it with DOOMED via vgonel() so that anyone who does find it
1938 	 * will skip over it.
1939 	 */
1940 	VI_LOCK(vp);
1941 	if (vp->v_usecount) {
1942 		VOP_UNLOCK(vp);
1943 		vdropl_recycle(vp);
1944 		vn_finished_write(vnmp);
1945 		CTR2(KTR_VFS,
1946 		    "%s: impossible to recycle, %p is already referenced",
1947 		    __func__, vp);
1948 		return (EBUSY);
1949 	}
1950 	if (!VN_IS_DOOMED(vp)) {
1951 		if (isvnlru)
1952 			recycles_free_count++;
1953 		else
1954 			counter_u64_add(direct_recycles_free_count, 1);
1955 		vgonel(vp);
1956 	}
1957 	VOP_UNLOCK(vp);
1958 	vdropl_recycle(vp);
1959 	vn_finished_write(vnmp);
1960 	return (0);
1961 }
1962 
1963 /*
1964  * Allocate a new vnode.
1965  *
1966  * The operation never returns an error. Returning an error was disabled
1967  * in r145385 (dated 2005) with the following comment:
1968  *
1969  * XXX Not all VFS_VGET/ffs_vget callers check returns.
1970  *
1971  * Given the age of this commit (almost 15 years at the time of writing this
1972  * comment) restoring the ability to fail requires a significant audit of
1973  * all codepaths.
1974  *
1975  * The routine can try to free a vnode or stall for up to 1 second waiting for
1976  * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1977  */
1978 static u_long vn_alloc_cyclecount;
1979 static u_long vn_alloc_sleeps;
1980 
1981 SYSCTL_ULONG(_vfs_vnode_stats, OID_AUTO, alloc_sleeps, CTLFLAG_RD, &vn_alloc_sleeps, 0,
1982     "Number of times vnode allocation blocked waiting on vnlru");
1983 
1984 static struct vnode * __noinline
vn_alloc_hard(struct mount * mp,u_long rnumvnodes,bool bumped)1985 vn_alloc_hard(struct mount *mp, u_long rnumvnodes, bool bumped)
1986 {
1987 	u_long rfreevnodes;
1988 
1989 	if (bumped) {
1990 		if (rnumvnodes > desiredvnodes + VNLRU_COUNT_SLOP) {
1991 			atomic_subtract_long(&numvnodes, 1);
1992 			bumped = false;
1993 		}
1994 	}
1995 
1996 	mtx_lock(&vnode_list_mtx);
1997 
1998 	rfreevnodes = vnlru_read_freevnodes();
1999 	if (vn_alloc_cyclecount++ >= rfreevnodes) {
2000 		vn_alloc_cyclecount = 0;
2001 		vstir = true;
2002 	}
2003 	/*
2004 	 * Grow the vnode cache if it will not be above its target max after
2005 	 * growing.  Otherwise, if there is at least one free vnode, try to
2006 	 * reclaim 1 item from it before growing the cache (possibly above its
2007 	 * target max if the reclamation failed or is delayed).
2008 	 */
2009 	if (vnlru_free_locked_direct(1) > 0)
2010 		goto alloc;
2011 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
2012 	if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
2013 		/*
2014 		 * Wait for space for a new vnode.
2015 		 */
2016 		if (bumped) {
2017 			atomic_subtract_long(&numvnodes, 1);
2018 			bumped = false;
2019 		}
2020 		mtx_lock(&vnode_list_mtx);
2021 		vnlru_kick_locked();
2022 		vn_alloc_sleeps++;
2023 		msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
2024 		if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
2025 		    vnlru_read_freevnodes() > 1)
2026 			vnlru_free_locked_direct(1);
2027 		else
2028 			mtx_unlock(&vnode_list_mtx);
2029 	}
2030 alloc:
2031 	mtx_assert(&vnode_list_mtx, MA_NOTOWNED);
2032 	if (!bumped)
2033 		atomic_add_long(&numvnodes, 1);
2034 	vnlru_kick_cond();
2035 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
2036 }
2037 
2038 static struct vnode *
vn_alloc(struct mount * mp)2039 vn_alloc(struct mount *mp)
2040 {
2041 	u_long rnumvnodes;
2042 
2043 	if (__predict_false(vn_alloc_cyclecount != 0))
2044 		return (vn_alloc_hard(mp, 0, false));
2045 	rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
2046 	if (__predict_false(vnlru_under(rnumvnodes, vlowat))) {
2047 		return (vn_alloc_hard(mp, rnumvnodes, true));
2048 	}
2049 
2050 	return (uma_zalloc_smr(vnode_zone, M_WAITOK));
2051 }
2052 
2053 static void
vn_free(struct vnode * vp)2054 vn_free(struct vnode *vp)
2055 {
2056 
2057 	atomic_subtract_long(&numvnodes, 1);
2058 	uma_zfree_smr(vnode_zone, vp);
2059 }
2060 
2061 /*
2062  * Allocate a new vnode.
2063  */
2064 int
getnewvnode(const char * tag,struct mount * mp,struct vop_vector * vops,struct vnode ** vpp)2065 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
2066     struct vnode **vpp)
2067 {
2068 	struct vnode *vp;
2069 	struct thread *td;
2070 	struct lock_object *lo;
2071 
2072 	CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
2073 
2074 	KASSERT(vops->registered,
2075 	    ("%s: not registered vector op %p\n", __func__, vops));
2076 	cache_validate_vop_vector(mp, vops);
2077 
2078 	td = curthread;
2079 	if (td->td_vp_reserved != NULL) {
2080 		vp = td->td_vp_reserved;
2081 		td->td_vp_reserved = NULL;
2082 	} else {
2083 		vp = vn_alloc(mp);
2084 	}
2085 	counter_u64_add(vnodes_created, 1);
2086 
2087 	vn_set_state(vp, VSTATE_UNINITIALIZED);
2088 
2089 	/*
2090 	 * Locks are given the generic name "vnode" when created.
2091 	 * Follow the historic practice of using the filesystem
2092 	 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
2093 	 *
2094 	 * Locks live in a witness group keyed on their name. Thus,
2095 	 * when a lock is renamed, it must also move from the witness
2096 	 * group of its old name to the witness group of its new name.
2097 	 *
2098 	 * The change only needs to be made when the vnode moves
2099 	 * from one filesystem type to another. We ensure that each
2100 	 * filesystem use a single static name pointer for its tag so
2101 	 * that we can compare pointers rather than doing a strcmp().
2102 	 */
2103 	lo = &vp->v_vnlock->lock_object;
2104 #ifdef WITNESS
2105 	if (lo->lo_name != tag) {
2106 #endif
2107 		lo->lo_name = tag;
2108 #ifdef WITNESS
2109 		WITNESS_DESTROY(lo);
2110 		WITNESS_INIT(lo, tag);
2111 	}
2112 #endif
2113 	/*
2114 	 * By default, don't allow shared locks unless filesystems opt-in.
2115 	 */
2116 	vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
2117 	/*
2118 	 * Finalize various vnode identity bits.
2119 	 */
2120 	KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
2121 	KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
2122 	KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
2123 	vp->v_type = VNON;
2124 	vp->v_op = vops;
2125 	vp->v_irflag = 0;
2126 	v_init_counters(vp);
2127 	vn_seqc_init(vp);
2128 	vp->v_bufobj.bo_ops = &buf_ops_bio;
2129 #ifdef DIAGNOSTIC
2130 	if (mp == NULL && vops != &dead_vnodeops)
2131 		printf("NULL mp in getnewvnode(9), tag %s\n", tag);
2132 #endif
2133 #ifdef MAC
2134 	mac_vnode_init(vp);
2135 	if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
2136 		mac_vnode_associate_singlelabel(mp, vp);
2137 #endif
2138 	if (mp != NULL) {
2139 		vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
2140 	}
2141 
2142 	/*
2143 	 * For the filesystems which do not use vfs_hash_insert(),
2144 	 * still initialize v_hash to have vfs_hash_index() useful.
2145 	 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
2146 	 * its own hashing.
2147 	 */
2148 	vp->v_hash = (uintptr_t)vp >> vnsz2log;
2149 
2150 	*vpp = vp;
2151 	return (0);
2152 }
2153 
2154 void
getnewvnode_reserve(void)2155 getnewvnode_reserve(void)
2156 {
2157 	struct thread *td;
2158 
2159 	td = curthread;
2160 	MPASS(td->td_vp_reserved == NULL);
2161 	td->td_vp_reserved = vn_alloc(NULL);
2162 }
2163 
2164 void
getnewvnode_drop_reserve(void)2165 getnewvnode_drop_reserve(void)
2166 {
2167 	struct thread *td;
2168 
2169 	td = curthread;
2170 	if (td->td_vp_reserved != NULL) {
2171 		vn_free(td->td_vp_reserved);
2172 		td->td_vp_reserved = NULL;
2173 	}
2174 }
2175 
2176 static void __noinline
freevnode(struct vnode * vp)2177 freevnode(struct vnode *vp)
2178 {
2179 	struct bufobj *bo;
2180 
2181 	/*
2182 	 * The vnode has been marked for destruction, so free it.
2183 	 *
2184 	 * The vnode will be returned to the zone where it will
2185 	 * normally remain until it is needed for another vnode. We
2186 	 * need to cleanup (or verify that the cleanup has already
2187 	 * been done) any residual data left from its current use
2188 	 * so as not to contaminate the freshly allocated vnode.
2189 	 */
2190 	CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
2191 	/*
2192 	 * Paired with vgone.
2193 	 */
2194 	vn_seqc_write_end_free(vp);
2195 
2196 	bo = &vp->v_bufobj;
2197 	VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
2198 	VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
2199 	VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
2200 	VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
2201 	VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
2202 	VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
2203 	VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
2204 	    ("clean blk trie not empty"));
2205 	VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
2206 	VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
2207 	    ("dirty blk trie not empty"));
2208 	VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
2209 	    ("Leaked inactivation"));
2210 	VI_UNLOCK(vp);
2211 	cache_assert_no_entries(vp);
2212 
2213 #ifdef MAC
2214 	mac_vnode_destroy(vp);
2215 #endif
2216 	if (vp->v_pollinfo != NULL) {
2217 		/*
2218 		 * Use LK_NOWAIT to shut up witness about the lock. We may get
2219 		 * here while having another vnode locked when trying to
2220 		 * satisfy a lookup and needing to recycle.
2221 		 */
2222 		VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT);
2223 		destroy_vpollinfo(vp->v_pollinfo);
2224 		VOP_UNLOCK(vp);
2225 		vp->v_pollinfo = NULL;
2226 	}
2227 	vp->v_mountedhere = NULL;
2228 	vp->v_unpcb = NULL;
2229 	vp->v_rdev = NULL;
2230 	vp->v_fifoinfo = NULL;
2231 	vp->v_iflag = 0;
2232 	vp->v_vflag = 0;
2233 	bo->bo_flag = 0;
2234 	vn_free(vp);
2235 }
2236 
2237 /*
2238  * Delete from old mount point vnode list, if on one.
2239  */
2240 static void
delmntque(struct vnode * vp)2241 delmntque(struct vnode *vp)
2242 {
2243 	struct mount *mp;
2244 
2245 	VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
2246 
2247 	mp = vp->v_mount;
2248 	MNT_ILOCK(mp);
2249 	VI_LOCK(vp);
2250 	vp->v_mount = NULL;
2251 	VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
2252 		("bad mount point vnode list size"));
2253 	TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2254 	mp->mnt_nvnodelistsize--;
2255 	MNT_REL(mp);
2256 	MNT_IUNLOCK(mp);
2257 	/*
2258 	 * The caller expects the interlock to be still held.
2259 	 */
2260 	ASSERT_VI_LOCKED(vp, __func__);
2261 }
2262 
2263 static int
insmntque1_int(struct vnode * vp,struct mount * mp,bool dtr)2264 insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr)
2265 {
2266 
2267 	KASSERT(vp->v_mount == NULL,
2268 		("insmntque: vnode already on per mount vnode list"));
2269 	VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
2270 	if ((mp->mnt_kern_flag & MNTK_UNLOCKED_INSMNTQUE) == 0) {
2271 		ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
2272 	} else {
2273 		KASSERT(!dtr,
2274 		    ("%s: can't have MNTK_UNLOCKED_INSMNTQUE and cleanup",
2275 		    __func__));
2276 	}
2277 
2278 	/*
2279 	 * We acquire the vnode interlock early to ensure that the
2280 	 * vnode cannot be recycled by another process releasing a
2281 	 * holdcnt on it before we get it on both the vnode list
2282 	 * and the active vnode list. The mount mutex protects only
2283 	 * manipulation of the vnode list and the vnode freelist
2284 	 * mutex protects only manipulation of the active vnode list.
2285 	 * Hence the need to hold the vnode interlock throughout.
2286 	 */
2287 	MNT_ILOCK(mp);
2288 	VI_LOCK(vp);
2289 	if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
2290 	    ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
2291 	    mp->mnt_nvnodelistsize == 0)) &&
2292 	    (vp->v_vflag & VV_FORCEINSMQ) == 0) {
2293 		VI_UNLOCK(vp);
2294 		MNT_IUNLOCK(mp);
2295 		if (dtr) {
2296 			vp->v_data = NULL;
2297 			vp->v_op = &dead_vnodeops;
2298 			vgone(vp);
2299 			vput(vp);
2300 		}
2301 		return (EBUSY);
2302 	}
2303 	vp->v_mount = mp;
2304 	MNT_REF(mp);
2305 	TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2306 	VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
2307 		("neg mount point vnode list size"));
2308 	mp->mnt_nvnodelistsize++;
2309 	VI_UNLOCK(vp);
2310 	MNT_IUNLOCK(mp);
2311 	return (0);
2312 }
2313 
2314 /*
2315  * Insert into list of vnodes for the new mount point, if available.
2316  * insmntque() reclaims the vnode on insertion failure, insmntque1()
2317  * leaves handling of the vnode to the caller.
2318  */
2319 int
insmntque(struct vnode * vp,struct mount * mp)2320 insmntque(struct vnode *vp, struct mount *mp)
2321 {
2322 	return (insmntque1_int(vp, mp, true));
2323 }
2324 
2325 int
insmntque1(struct vnode * vp,struct mount * mp)2326 insmntque1(struct vnode *vp, struct mount *mp)
2327 {
2328 	return (insmntque1_int(vp, mp, false));
2329 }
2330 
2331 /*
2332  * Flush out and invalidate all buffers associated with a bufobj
2333  * Called with the underlying object locked.
2334  */
2335 int
bufobj_invalbuf(struct bufobj * bo,int flags,int slpflag,int slptimeo)2336 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2337 {
2338 	int error;
2339 
2340 	BO_LOCK(bo);
2341 	if (flags & V_SAVE) {
2342 		error = bufobj_wwait(bo, slpflag, slptimeo);
2343 		if (error) {
2344 			BO_UNLOCK(bo);
2345 			return (error);
2346 		}
2347 		if (bo->bo_dirty.bv_cnt > 0) {
2348 			BO_UNLOCK(bo);
2349 			do {
2350 				error = BO_SYNC(bo, MNT_WAIT);
2351 			} while (error == ERELOOKUP);
2352 			if (error != 0)
2353 				return (error);
2354 			BO_LOCK(bo);
2355 			if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2356 				BO_UNLOCK(bo);
2357 				return (EBUSY);
2358 			}
2359 		}
2360 	}
2361 	/*
2362 	 * If you alter this loop please notice that interlock is dropped and
2363 	 * reacquired in flushbuflist.  Special care is needed to ensure that
2364 	 * no race conditions occur from this.
2365 	 */
2366 	do {
2367 		error = flushbuflist(&bo->bo_clean,
2368 		    flags, bo, slpflag, slptimeo);
2369 		if (error == 0 && !(flags & V_CLEANONLY))
2370 			error = flushbuflist(&bo->bo_dirty,
2371 			    flags, bo, slpflag, slptimeo);
2372 		if (error != 0 && error != EAGAIN) {
2373 			BO_UNLOCK(bo);
2374 			return (error);
2375 		}
2376 	} while (error != 0);
2377 
2378 	/*
2379 	 * Wait for I/O to complete.  XXX needs cleaning up.  The vnode can
2380 	 * have write I/O in-progress but if there is a VM object then the
2381 	 * VM object can also have read-I/O in-progress.
2382 	 */
2383 	do {
2384 		bufobj_wwait(bo, 0, 0);
2385 		if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2386 			BO_UNLOCK(bo);
2387 			vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2388 			BO_LOCK(bo);
2389 		}
2390 	} while (bo->bo_numoutput > 0);
2391 	BO_UNLOCK(bo);
2392 
2393 	/*
2394 	 * Destroy the copy in the VM cache, too.
2395 	 */
2396 	if (bo->bo_object != NULL &&
2397 	    (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2398 		VM_OBJECT_WLOCK(bo->bo_object);
2399 		vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2400 		    OBJPR_CLEANONLY : 0);
2401 		VM_OBJECT_WUNLOCK(bo->bo_object);
2402 	}
2403 
2404 #ifdef INVARIANTS
2405 	BO_LOCK(bo);
2406 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2407 	    V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2408 	    bo->bo_clean.bv_cnt > 0))
2409 		panic("vinvalbuf: flush failed");
2410 	if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2411 	    bo->bo_dirty.bv_cnt > 0)
2412 		panic("vinvalbuf: flush dirty failed");
2413 	BO_UNLOCK(bo);
2414 #endif
2415 	return (0);
2416 }
2417 
2418 /*
2419  * Flush out and invalidate all buffers associated with a vnode.
2420  * Called with the underlying object locked.
2421  */
2422 int
vinvalbuf(struct vnode * vp,int flags,int slpflag,int slptimeo)2423 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2424 {
2425 
2426 	CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2427 	ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2428 	if (vp->v_object != NULL && vp->v_object->handle != vp)
2429 		return (0);
2430 	return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2431 }
2432 
2433 /*
2434  * Flush out buffers on the specified list.
2435  *
2436  */
2437 static int
flushbuflist(struct bufv * bufv,int flags,struct bufobj * bo,int slpflag,int slptimeo)2438 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2439     int slptimeo)
2440 {
2441 	struct buf *bp, *nbp;
2442 	int retval, error;
2443 	daddr_t lblkno;
2444 	b_xflags_t xflags;
2445 
2446 	ASSERT_BO_WLOCKED(bo);
2447 
2448 	retval = 0;
2449 	TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2450 		/*
2451 		 * If we are flushing both V_NORMAL and V_ALT buffers then
2452 		 * do not skip any buffers. If we are flushing only V_NORMAL
2453 		 * buffers then skip buffers marked as BX_ALTDATA. If we are
2454 		 * flushing only V_ALT buffers then skip buffers not marked
2455 		 * as BX_ALTDATA.
2456 		 */
2457 		if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2458 		   (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2459 		    ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2460 			continue;
2461 		}
2462 		if (nbp != NULL) {
2463 			lblkno = nbp->b_lblkno;
2464 			xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2465 		}
2466 		retval = EAGAIN;
2467 		error = BUF_TIMELOCK(bp,
2468 		    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2469 		    "flushbuf", slpflag, slptimeo);
2470 		if (error) {
2471 			BO_LOCK(bo);
2472 			return (error != ENOLCK ? error : EAGAIN);
2473 		}
2474 		KASSERT(bp->b_bufobj == bo,
2475 		    ("bp %p wrong b_bufobj %p should be %p",
2476 		    bp, bp->b_bufobj, bo));
2477 		/*
2478 		 * XXX Since there are no node locks for NFS, I
2479 		 * believe there is a slight chance that a delayed
2480 		 * write will occur while sleeping just above, so
2481 		 * check for it.
2482 		 */
2483 		if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2484 		    (flags & V_SAVE)) {
2485 			bremfree(bp);
2486 			bp->b_flags |= B_ASYNC;
2487 			bwrite(bp);
2488 			BO_LOCK(bo);
2489 			return (EAGAIN);	/* XXX: why not loop ? */
2490 		}
2491 		bremfree(bp);
2492 		bp->b_flags |= (B_INVAL | B_RELBUF);
2493 		bp->b_flags &= ~B_ASYNC;
2494 		brelse(bp);
2495 		BO_LOCK(bo);
2496 		if (nbp == NULL)
2497 			break;
2498 		nbp = gbincore(bo, lblkno);
2499 		if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2500 		    != xflags)
2501 			break;			/* nbp invalid */
2502 	}
2503 	return (retval);
2504 }
2505 
2506 int
bnoreuselist(struct bufv * bufv,struct bufobj * bo,daddr_t startn,daddr_t endn)2507 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2508 {
2509 	struct buf *bp;
2510 	int error;
2511 	daddr_t lblkno;
2512 
2513 	ASSERT_BO_LOCKED(bo);
2514 
2515 	for (lblkno = startn;;) {
2516 again:
2517 		bp = buf_lookup_ge(bufv, lblkno);
2518 		if (bp == NULL || bp->b_lblkno >= endn)
2519 			break;
2520 		error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2521 		    LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2522 		if (error != 0) {
2523 			BO_RLOCK(bo);
2524 			if (error == ENOLCK)
2525 				goto again;
2526 			return (error);
2527 		}
2528 		KASSERT(bp->b_bufobj == bo,
2529 		    ("bp %p wrong b_bufobj %p should be %p",
2530 		    bp, bp->b_bufobj, bo));
2531 		lblkno = bp->b_lblkno + 1;
2532 		if ((bp->b_flags & B_MANAGED) == 0)
2533 			bremfree(bp);
2534 		bp->b_flags |= B_RELBUF;
2535 		/*
2536 		 * In the VMIO case, use the B_NOREUSE flag to hint that the
2537 		 * pages backing each buffer in the range are unlikely to be
2538 		 * reused.  Dirty buffers will have the hint applied once
2539 		 * they've been written.
2540 		 */
2541 		if ((bp->b_flags & B_VMIO) != 0)
2542 			bp->b_flags |= B_NOREUSE;
2543 		brelse(bp);
2544 		BO_RLOCK(bo);
2545 	}
2546 	return (0);
2547 }
2548 
2549 /*
2550  * Truncate a file's buffer and pages to a specified length.  This
2551  * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2552  * sync activity.
2553  */
2554 int
vtruncbuf(struct vnode * vp,off_t length,int blksize)2555 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2556 {
2557 	struct buf *bp, *nbp;
2558 	struct bufobj *bo;
2559 	daddr_t startlbn;
2560 
2561 	CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2562 	    vp, blksize, (uintmax_t)length);
2563 
2564 	/*
2565 	 * Round up to the *next* lbn.
2566 	 */
2567 	startlbn = howmany(length, blksize);
2568 
2569 	ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2570 
2571 	bo = &vp->v_bufobj;
2572 restart_unlocked:
2573 	BO_LOCK(bo);
2574 
2575 	while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2576 		;
2577 
2578 	if (length > 0) {
2579 		/*
2580 		 * Write out vnode metadata, e.g. indirect blocks.
2581 		 */
2582 restartsync:
2583 		TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2584 			if (bp->b_lblkno >= 0)
2585 				continue;
2586 			/*
2587 			 * Since we hold the vnode lock this should only
2588 			 * fail if we're racing with the buf daemon.
2589 			 */
2590 			if (BUF_LOCK(bp,
2591 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2592 			    BO_LOCKPTR(bo)) == ENOLCK)
2593 				goto restart_unlocked;
2594 
2595 			VNASSERT((bp->b_flags & B_DELWRI), vp,
2596 			    ("buf(%p) on dirty queue without DELWRI", bp));
2597 
2598 			bremfree(bp);
2599 			bawrite(bp);
2600 			BO_LOCK(bo);
2601 			goto restartsync;
2602 		}
2603 	}
2604 
2605 	bufobj_wwait(bo, 0, 0);
2606 	BO_UNLOCK(bo);
2607 	vnode_pager_setsize(vp, length);
2608 
2609 	return (0);
2610 }
2611 
2612 /*
2613  * Invalidate the cached pages of a file's buffer within the range of block
2614  * numbers [startlbn, endlbn).
2615  */
2616 void
v_inval_buf_range(struct vnode * vp,daddr_t startlbn,daddr_t endlbn,int blksize)2617 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2618     int blksize)
2619 {
2620 	struct bufobj *bo;
2621 	off_t start, end;
2622 
2623 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2624 
2625 	start = blksize * startlbn;
2626 	end = blksize * endlbn;
2627 
2628 	bo = &vp->v_bufobj;
2629 	BO_LOCK(bo);
2630 	MPASS(blksize == bo->bo_bsize);
2631 
2632 	while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2633 		;
2634 
2635 	BO_UNLOCK(bo);
2636 	vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2637 }
2638 
2639 static int
v_inval_buf_range_locked(struct vnode * vp,struct bufobj * bo,daddr_t startlbn,daddr_t endlbn)2640 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2641     daddr_t startlbn, daddr_t endlbn)
2642 {
2643 	struct bufv *bv;
2644 	struct buf *bp, *nbp;
2645 	uint8_t anyfreed;
2646 	bool clean;
2647 
2648 	ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2649 	ASSERT_BO_LOCKED(bo);
2650 
2651 	anyfreed = 1;
2652 	clean = true;
2653 	do {
2654 		bv = clean ? &bo->bo_clean : &bo->bo_dirty;
2655 		bp = buf_lookup_ge(bv, startlbn);
2656 		if (bp == NULL)
2657 			continue;
2658 		TAILQ_FOREACH_FROM_SAFE(bp, &bv->bv_hd, b_bobufs, nbp) {
2659 			if (bp->b_lblkno >= endlbn)
2660 				break;
2661 			if (BUF_LOCK(bp,
2662 			    LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2663 			    BO_LOCKPTR(bo)) == ENOLCK) {
2664 				BO_LOCK(bo);
2665 				return (EAGAIN);
2666 			}
2667 
2668 			bremfree(bp);
2669 			bp->b_flags |= B_INVAL | B_RELBUF;
2670 			bp->b_flags &= ~B_ASYNC;
2671 			brelse(bp);
2672 			anyfreed = 2;
2673 
2674 			BO_LOCK(bo);
2675 			if (nbp != NULL &&
2676 			    (((nbp->b_xflags &
2677 			    (clean ? BX_VNCLEAN : BX_VNDIRTY)) == 0) ||
2678 			    nbp->b_vp != vp ||
2679 			    (nbp->b_flags & B_DELWRI) == (clean? B_DELWRI: 0)))
2680 				return (EAGAIN);
2681 		}
2682 	} while (clean = !clean, anyfreed-- > 0);
2683 	return (0);
2684 }
2685 
2686 static void
buf_vlist_remove(struct buf * bp)2687 buf_vlist_remove(struct buf *bp)
2688 {
2689 	struct bufv *bv;
2690 	b_xflags_t flags;
2691 
2692 	flags = bp->b_xflags;
2693 
2694 	KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2695 	ASSERT_BO_WLOCKED(bp->b_bufobj);
2696 	KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2697 	    (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2698 	    ("%s: buffer %p has invalid queue state", __func__, bp));
2699 
2700 	if ((flags & BX_VNDIRTY) != 0)
2701 		bv = &bp->b_bufobj->bo_dirty;
2702 	else
2703 		bv = &bp->b_bufobj->bo_clean;
2704 	BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2705 	TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2706 	bv->bv_cnt--;
2707 	bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2708 }
2709 
2710 /*
2711  * Add the buffer to the sorted clean or dirty block list.  Return zero on
2712  * success, EEXIST if a buffer with this identity already exists, or another
2713  * error on allocation failure.
2714  */
2715 static inline int
buf_vlist_find_or_add(struct buf * bp,struct bufobj * bo,b_xflags_t xflags)2716 buf_vlist_find_or_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2717 {
2718 	struct bufv *bv;
2719 	struct buf *n;
2720 	int error;
2721 
2722 	ASSERT_BO_WLOCKED(bo);
2723 	KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2724 	    ("buf_vlist_add: bo %p does not allow bufs", bo));
2725 	KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2726 	    ("dead bo %p", bo));
2727 	KASSERT((bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) == xflags,
2728 	    ("buf_vlist_add: b_xflags %#x not set on bp %p", xflags, bp));
2729 
2730 	if (xflags & BX_VNDIRTY)
2731 		bv = &bo->bo_dirty;
2732 	else
2733 		bv = &bo->bo_clean;
2734 
2735 	error = buf_insert_lookup_le(bv, bp, &n);
2736 	if (n == NULL) {
2737 		KASSERT(error != EEXIST,
2738 		    ("buf_vlist_add: EEXIST but no existing buf found: bp %p",
2739 		    bp));
2740 	} else {
2741 		KASSERT(n->b_lblkno <= bp->b_lblkno,
2742 		    ("buf_vlist_add: out of order insert/lookup: bp %p n %p",
2743 		    bp, n));
2744 		KASSERT((n->b_lblkno == bp->b_lblkno) == (error == EEXIST),
2745 		    ("buf_vlist_add: inconsistent result for existing buf: "
2746 		    "error %d bp %p n %p", error, bp, n));
2747 	}
2748 	if (error != 0)
2749 		return (error);
2750 
2751 	/* Keep the list ordered. */
2752 	if (n == NULL) {
2753 		KASSERT(TAILQ_EMPTY(&bv->bv_hd) ||
2754 		    bp->b_lblkno < TAILQ_FIRST(&bv->bv_hd)->b_lblkno,
2755 		    ("buf_vlist_add: queue order: "
2756 		    "%p should be before first %p",
2757 		    bp, TAILQ_FIRST(&bv->bv_hd)));
2758 		TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2759 	} else {
2760 		KASSERT(TAILQ_NEXT(n, b_bobufs) == NULL ||
2761 		    bp->b_lblkno < TAILQ_NEXT(n, b_bobufs)->b_lblkno,
2762 		    ("buf_vlist_add: queue order: "
2763 		    "%p should be before next %p",
2764 		    bp, TAILQ_NEXT(n, b_bobufs)));
2765 		TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2766 	}
2767 
2768 	bv->bv_cnt++;
2769 	return (0);
2770 }
2771 
2772 /*
2773  * Add the buffer to the sorted clean or dirty block list.
2774  *
2775  * NOTE: xflags is passed as a constant, optimizing this inline function!
2776  */
2777 static void
buf_vlist_add(struct buf * bp,struct bufobj * bo,b_xflags_t xflags)2778 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2779 {
2780 	int error;
2781 
2782 	KASSERT((bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN)) == 0,
2783 	    ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2784 	bp->b_xflags |= xflags;
2785 	error = buf_vlist_find_or_add(bp, bo, xflags);
2786 	if (error)
2787 		panic("buf_vlist_add: error=%d", error);
2788 }
2789 
2790 /*
2791  * Look up a buffer using the buffer tries.
2792  */
2793 struct buf *
gbincore(struct bufobj * bo,daddr_t lblkno)2794 gbincore(struct bufobj *bo, daddr_t lblkno)
2795 {
2796 	struct buf *bp;
2797 
2798 	ASSERT_BO_LOCKED(bo);
2799 	bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2800 	if (bp != NULL)
2801 		return (bp);
2802 	return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2803 }
2804 
2805 /*
2806  * Look up a buf using the buffer tries, without the bufobj lock.  This relies
2807  * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2808  * stability of the result.  Like other lockless lookups, the found buf may
2809  * already be invalid by the time this function returns.
2810  */
2811 struct buf *
gbincore_unlocked(struct bufobj * bo,daddr_t lblkno)2812 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2813 {
2814 	struct buf *bp;
2815 
2816 	ASSERT_BO_UNLOCKED(bo);
2817 	bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2818 	if (bp != NULL)
2819 		return (bp);
2820 	return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2821 }
2822 
2823 /*
2824  * Associate a buffer with a vnode.
2825  */
2826 int
bgetvp(struct vnode * vp,struct buf * bp)2827 bgetvp(struct vnode *vp, struct buf *bp)
2828 {
2829 	struct bufobj *bo;
2830 	int error;
2831 
2832 	bo = &vp->v_bufobj;
2833 	ASSERT_BO_UNLOCKED(bo);
2834 	VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2835 
2836 	CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2837 	VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2838 	    ("bgetvp: bp already attached! %p", bp));
2839 
2840 	/*
2841 	 * Add the buf to the vnode's clean list unless we lost a race and find
2842 	 * an existing buf in either dirty or clean.
2843 	 */
2844 	bp->b_vp = vp;
2845 	bp->b_bufobj = bo;
2846 	bp->b_xflags |= BX_VNCLEAN;
2847 	error = EEXIST;
2848 	BO_LOCK(bo);
2849 	if (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, bp->b_lblkno) == NULL)
2850 		error = buf_vlist_find_or_add(bp, bo, BX_VNCLEAN);
2851 	BO_UNLOCK(bo);
2852 	if (__predict_true(error == 0)) {
2853 		vhold(vp);
2854 		return (0);
2855 	}
2856 	if (error != EEXIST)
2857 		panic("bgetvp: buf_vlist_add error: %d", error);
2858 	bp->b_vp = NULL;
2859 	bp->b_bufobj = NULL;
2860 	bp->b_xflags &= ~BX_VNCLEAN;
2861 	return (error);
2862 }
2863 
2864 /*
2865  * Disassociate a buffer from a vnode.
2866  */
2867 void
brelvp(struct buf * bp)2868 brelvp(struct buf *bp)
2869 {
2870 	struct bufobj *bo;
2871 	struct vnode *vp;
2872 
2873 	CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2874 	KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2875 
2876 	/*
2877 	 * Delete from old vnode list, if on one.
2878 	 */
2879 	vp = bp->b_vp;		/* XXX */
2880 	bo = bp->b_bufobj;
2881 	BO_LOCK(bo);
2882 	buf_vlist_remove(bp);
2883 	if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2884 		bo->bo_flag &= ~BO_ONWORKLST;
2885 		mtx_lock(&sync_mtx);
2886 		LIST_REMOVE(bo, bo_synclist);
2887 		syncer_worklist_len--;
2888 		mtx_unlock(&sync_mtx);
2889 	}
2890 	bp->b_vp = NULL;
2891 	bp->b_bufobj = NULL;
2892 	BO_UNLOCK(bo);
2893 	vdrop(vp);
2894 }
2895 
2896 /*
2897  * Add an item to the syncer work queue.
2898  */
2899 static void
vn_syncer_add_to_worklist(struct bufobj * bo,int delay)2900 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2901 {
2902 	int slot;
2903 
2904 	ASSERT_BO_WLOCKED(bo);
2905 
2906 	mtx_lock(&sync_mtx);
2907 	if (bo->bo_flag & BO_ONWORKLST)
2908 		LIST_REMOVE(bo, bo_synclist);
2909 	else {
2910 		bo->bo_flag |= BO_ONWORKLST;
2911 		syncer_worklist_len++;
2912 	}
2913 
2914 	if (delay > syncer_maxdelay - 2)
2915 		delay = syncer_maxdelay - 2;
2916 	slot = (syncer_delayno + delay) & syncer_mask;
2917 
2918 	LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2919 	mtx_unlock(&sync_mtx);
2920 }
2921 
2922 static int
sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)2923 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2924 {
2925 	int error, len;
2926 
2927 	mtx_lock(&sync_mtx);
2928 	len = syncer_worklist_len - sync_vnode_count;
2929 	mtx_unlock(&sync_mtx);
2930 	error = SYSCTL_OUT(req, &len, sizeof(len));
2931 	return (error);
2932 }
2933 
2934 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2935     CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2936     sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2937 
2938 static struct proc *updateproc;
2939 static void sched_sync(void);
2940 static struct kproc_desc up_kp = {
2941 	"syncer",
2942 	sched_sync,
2943 	&updateproc
2944 };
2945 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2946 
2947 static int
sync_vnode(struct synclist * slp,struct bufobj ** bo,struct thread * td)2948 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2949 {
2950 	struct vnode *vp;
2951 	struct mount *mp;
2952 
2953 	*bo = LIST_FIRST(slp);
2954 	if (*bo == NULL)
2955 		return (0);
2956 	vp = bo2vnode(*bo);
2957 	if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2958 		return (1);
2959 	/*
2960 	 * We use vhold in case the vnode does not
2961 	 * successfully sync.  vhold prevents the vnode from
2962 	 * going away when we unlock the sync_mtx so that
2963 	 * we can acquire the vnode interlock.
2964 	 */
2965 	vholdl(vp);
2966 	mtx_unlock(&sync_mtx);
2967 	VI_UNLOCK(vp);
2968 	if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2969 		vdrop(vp);
2970 		mtx_lock(&sync_mtx);
2971 		return (*bo == LIST_FIRST(slp));
2972 	}
2973 	MPASSERT(mp == NULL || (curthread->td_pflags & TDP_IGNSUSP) != 0 ||
2974 	    (mp->mnt_kern_flag & MNTK_SUSPENDED) == 0, mp,
2975 	    ("suspended mp syncing vp %p", vp));
2976 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2977 	(void) VOP_FSYNC(vp, MNT_LAZY, td);
2978 	VOP_UNLOCK(vp);
2979 	vn_finished_write(mp);
2980 	BO_LOCK(*bo);
2981 	if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2982 		/*
2983 		 * Put us back on the worklist.  The worklist
2984 		 * routine will remove us from our current
2985 		 * position and then add us back in at a later
2986 		 * position.
2987 		 */
2988 		vn_syncer_add_to_worklist(*bo, syncdelay);
2989 	}
2990 	BO_UNLOCK(*bo);
2991 	vdrop(vp);
2992 	mtx_lock(&sync_mtx);
2993 	return (0);
2994 }
2995 
2996 static int first_printf = 1;
2997 
2998 /*
2999  * System filesystem synchronizer daemon.
3000  */
3001 static void
sched_sync(void)3002 sched_sync(void)
3003 {
3004 	struct synclist *next, *slp;
3005 	struct bufobj *bo;
3006 	long starttime;
3007 	struct thread *td = curthread;
3008 	int last_work_seen;
3009 	int net_worklist_len;
3010 	int syncer_final_iter;
3011 	int error;
3012 
3013 	last_work_seen = 0;
3014 	syncer_final_iter = 0;
3015 	syncer_state = SYNCER_RUNNING;
3016 	starttime = time_uptime;
3017 	td->td_pflags |= TDP_NORUNNINGBUF;
3018 
3019 	EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
3020 	    SHUTDOWN_PRI_LAST);
3021 
3022 	mtx_lock(&sync_mtx);
3023 	for (;;) {
3024 		if (syncer_state == SYNCER_FINAL_DELAY &&
3025 		    syncer_final_iter == 0) {
3026 			mtx_unlock(&sync_mtx);
3027 			kproc_suspend_check(td->td_proc);
3028 			mtx_lock(&sync_mtx);
3029 		}
3030 		net_worklist_len = syncer_worklist_len - sync_vnode_count;
3031 		if (syncer_state != SYNCER_RUNNING &&
3032 		    starttime != time_uptime) {
3033 			if (first_printf) {
3034 				printf("\nSyncing disks, vnodes remaining... ");
3035 				first_printf = 0;
3036 			}
3037 			printf("%d ", net_worklist_len);
3038 		}
3039 		starttime = time_uptime;
3040 
3041 		/*
3042 		 * Push files whose dirty time has expired.  Be careful
3043 		 * of interrupt race on slp queue.
3044 		 *
3045 		 * Skip over empty worklist slots when shutting down.
3046 		 */
3047 		do {
3048 			slp = &syncer_workitem_pending[syncer_delayno];
3049 			syncer_delayno += 1;
3050 			if (syncer_delayno == syncer_maxdelay)
3051 				syncer_delayno = 0;
3052 			next = &syncer_workitem_pending[syncer_delayno];
3053 			/*
3054 			 * If the worklist has wrapped since the
3055 			 * it was emptied of all but syncer vnodes,
3056 			 * switch to the FINAL_DELAY state and run
3057 			 * for one more second.
3058 			 */
3059 			if (syncer_state == SYNCER_SHUTTING_DOWN &&
3060 			    net_worklist_len == 0 &&
3061 			    last_work_seen == syncer_delayno) {
3062 				syncer_state = SYNCER_FINAL_DELAY;
3063 				syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
3064 			}
3065 		} while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
3066 		    syncer_worklist_len > 0);
3067 
3068 		/*
3069 		 * Keep track of the last time there was anything
3070 		 * on the worklist other than syncer vnodes.
3071 		 * Return to the SHUTTING_DOWN state if any
3072 		 * new work appears.
3073 		 */
3074 		if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
3075 			last_work_seen = syncer_delayno;
3076 		if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
3077 			syncer_state = SYNCER_SHUTTING_DOWN;
3078 		while (!LIST_EMPTY(slp)) {
3079 			error = sync_vnode(slp, &bo, td);
3080 			if (error == 1) {
3081 				LIST_REMOVE(bo, bo_synclist);
3082 				LIST_INSERT_HEAD(next, bo, bo_synclist);
3083 				continue;
3084 			}
3085 
3086 			if (first_printf == 0) {
3087 				/*
3088 				 * Drop the sync mutex, because some watchdog
3089 				 * drivers need to sleep while patting
3090 				 */
3091 				mtx_unlock(&sync_mtx);
3092 				wdog_kern_pat(WD_LASTVAL);
3093 				mtx_lock(&sync_mtx);
3094 			}
3095 		}
3096 		if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
3097 			syncer_final_iter--;
3098 		/*
3099 		 * The variable rushjob allows the kernel to speed up the
3100 		 * processing of the filesystem syncer process. A rushjob
3101 		 * value of N tells the filesystem syncer to process the next
3102 		 * N seconds worth of work on its queue ASAP. Currently rushjob
3103 		 * is used by the soft update code to speed up the filesystem
3104 		 * syncer process when the incore state is getting so far
3105 		 * ahead of the disk that the kernel memory pool is being
3106 		 * threatened with exhaustion.
3107 		 */
3108 		if (rushjob > 0) {
3109 			rushjob -= 1;
3110 			continue;
3111 		}
3112 		/*
3113 		 * Just sleep for a short period of time between
3114 		 * iterations when shutting down to allow some I/O
3115 		 * to happen.
3116 		 *
3117 		 * If it has taken us less than a second to process the
3118 		 * current work, then wait. Otherwise start right over
3119 		 * again. We can still lose time if any single round
3120 		 * takes more than two seconds, but it does not really
3121 		 * matter as we are just trying to generally pace the
3122 		 * filesystem activity.
3123 		 */
3124 		if (syncer_state != SYNCER_RUNNING ||
3125 		    time_uptime == starttime) {
3126 			thread_lock(td);
3127 			sched_prio(td, PPAUSE);
3128 			thread_unlock(td);
3129 		}
3130 		if (syncer_state != SYNCER_RUNNING)
3131 			cv_timedwait(&sync_wakeup, &sync_mtx,
3132 			    hz / SYNCER_SHUTDOWN_SPEEDUP);
3133 		else if (time_uptime == starttime)
3134 			cv_timedwait(&sync_wakeup, &sync_mtx, hz);
3135 	}
3136 }
3137 
3138 /*
3139  * Request the syncer daemon to speed up its work.
3140  * We never push it to speed up more than half of its
3141  * normal turn time, otherwise it could take over the cpu.
3142  */
3143 int
speedup_syncer(void)3144 speedup_syncer(void)
3145 {
3146 	int ret = 0;
3147 
3148 	mtx_lock(&sync_mtx);
3149 	if (rushjob < syncdelay / 2) {
3150 		rushjob += 1;
3151 		stat_rush_requests += 1;
3152 		ret = 1;
3153 	}
3154 	mtx_unlock(&sync_mtx);
3155 	cv_broadcast(&sync_wakeup);
3156 	return (ret);
3157 }
3158 
3159 /*
3160  * Tell the syncer to speed up its work and run though its work
3161  * list several times, then tell it to shut down.
3162  */
3163 static void
syncer_shutdown(void * arg,int howto)3164 syncer_shutdown(void *arg, int howto)
3165 {
3166 
3167 	if (howto & RB_NOSYNC)
3168 		return;
3169 	mtx_lock(&sync_mtx);
3170 	syncer_state = SYNCER_SHUTTING_DOWN;
3171 	rushjob = 0;
3172 	mtx_unlock(&sync_mtx);
3173 	cv_broadcast(&sync_wakeup);
3174 	kproc_shutdown(arg, howto);
3175 }
3176 
3177 void
syncer_suspend(void)3178 syncer_suspend(void)
3179 {
3180 
3181 	syncer_shutdown(updateproc, 0);
3182 }
3183 
3184 void
syncer_resume(void)3185 syncer_resume(void)
3186 {
3187 
3188 	mtx_lock(&sync_mtx);
3189 	first_printf = 1;
3190 	syncer_state = SYNCER_RUNNING;
3191 	mtx_unlock(&sync_mtx);
3192 	cv_broadcast(&sync_wakeup);
3193 	kproc_resume(updateproc);
3194 }
3195 
3196 /*
3197  * Move the buffer between the clean and dirty lists of its vnode.
3198  */
3199 void
reassignbuf(struct buf * bp)3200 reassignbuf(struct buf *bp)
3201 {
3202 	struct vnode *vp;
3203 	struct bufobj *bo;
3204 	int delay;
3205 #ifdef INVARIANTS
3206 	struct bufv *bv;
3207 #endif
3208 
3209 	vp = bp->b_vp;
3210 	bo = bp->b_bufobj;
3211 
3212 	KASSERT((bp->b_flags & B_PAGING) == 0,
3213 	    ("%s: cannot reassign paging buffer %p", __func__, bp));
3214 
3215 	CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
3216 	    bp, bp->b_vp, bp->b_flags);
3217 
3218 	BO_LOCK(bo);
3219 	if ((bo->bo_flag & BO_NONSTERILE) == 0) {
3220 		/*
3221 		 * Coordinate with getblk's unlocked lookup.  Make
3222 		 * BO_NONSTERILE visible before the first reassignbuf produces
3223 		 * any side effect.  This could be outside the bo lock if we
3224 		 * used a separate atomic flag field.
3225 		 */
3226 		bo->bo_flag |= BO_NONSTERILE;
3227 		atomic_thread_fence_rel();
3228 	}
3229 	buf_vlist_remove(bp);
3230 
3231 	/*
3232 	 * If dirty, put on list of dirty buffers; otherwise insert onto list
3233 	 * of clean buffers.
3234 	 */
3235 	if (bp->b_flags & B_DELWRI) {
3236 		if ((bo->bo_flag & BO_ONWORKLST) == 0) {
3237 			switch (vp->v_type) {
3238 			case VDIR:
3239 				delay = dirdelay;
3240 				break;
3241 			case VCHR:
3242 				delay = metadelay;
3243 				break;
3244 			default:
3245 				delay = filedelay;
3246 			}
3247 			vn_syncer_add_to_worklist(bo, delay);
3248 		}
3249 		buf_vlist_add(bp, bo, BX_VNDIRTY);
3250 	} else {
3251 		buf_vlist_add(bp, bo, BX_VNCLEAN);
3252 
3253 		if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
3254 			mtx_lock(&sync_mtx);
3255 			LIST_REMOVE(bo, bo_synclist);
3256 			syncer_worklist_len--;
3257 			mtx_unlock(&sync_mtx);
3258 			bo->bo_flag &= ~BO_ONWORKLST;
3259 		}
3260 	}
3261 #ifdef INVARIANTS
3262 	bv = &bo->bo_clean;
3263 	bp = TAILQ_FIRST(&bv->bv_hd);
3264 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3265 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3266 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
3267 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3268 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3269 	bv = &bo->bo_dirty;
3270 	bp = TAILQ_FIRST(&bv->bv_hd);
3271 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3272 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3273 	bp = TAILQ_LAST(&bv->bv_hd, buflists);
3274 	KASSERT(bp == NULL || bp->b_bufobj == bo,
3275 	    ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
3276 #endif
3277 	BO_UNLOCK(bo);
3278 }
3279 
3280 static void
v_init_counters(struct vnode * vp)3281 v_init_counters(struct vnode *vp)
3282 {
3283 
3284 	VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
3285 	    vp, ("%s called for an initialized vnode", __FUNCTION__));
3286 	ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
3287 
3288 	refcount_init(&vp->v_holdcnt, 1);
3289 	refcount_init(&vp->v_usecount, 1);
3290 }
3291 
3292 /*
3293  * Get a usecount on a vnode.
3294  *
3295  * vget and vget_finish may fail to lock the vnode if they lose a race against
3296  * it being doomed. LK_RETRY can be passed in flags to lock it anyway.
3297  *
3298  * Consumers which don't guarantee liveness of the vnode can use SMR to
3299  * try to get a reference. Note this operation can fail since the vnode
3300  * may be awaiting getting freed by the time they get to it.
3301  */
3302 enum vgetstate
vget_prep_smr(struct vnode * vp)3303 vget_prep_smr(struct vnode *vp)
3304 {
3305 	enum vgetstate vs;
3306 
3307 	VFS_SMR_ASSERT_ENTERED();
3308 
3309 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3310 		vs = VGET_USECOUNT;
3311 	} else {
3312 		if (vhold_smr(vp))
3313 			vs = VGET_HOLDCNT;
3314 		else
3315 			vs = VGET_NONE;
3316 	}
3317 	return (vs);
3318 }
3319 
3320 enum vgetstate
vget_prep(struct vnode * vp)3321 vget_prep(struct vnode *vp)
3322 {
3323 	enum vgetstate vs;
3324 
3325 	if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
3326 		vs = VGET_USECOUNT;
3327 	} else {
3328 		vhold(vp);
3329 		vs = VGET_HOLDCNT;
3330 	}
3331 	return (vs);
3332 }
3333 
3334 void
vget_abort(struct vnode * vp,enum vgetstate vs)3335 vget_abort(struct vnode *vp, enum vgetstate vs)
3336 {
3337 
3338 	switch (vs) {
3339 	case VGET_USECOUNT:
3340 		vrele(vp);
3341 		break;
3342 	case VGET_HOLDCNT:
3343 		vdrop(vp);
3344 		break;
3345 	default:
3346 		__assert_unreachable();
3347 	}
3348 }
3349 
3350 int
vget(struct vnode * vp,int flags)3351 vget(struct vnode *vp, int flags)
3352 {
3353 	enum vgetstate vs;
3354 
3355 	vs = vget_prep(vp);
3356 	return (vget_finish(vp, flags, vs));
3357 }
3358 
3359 int
vget_finish(struct vnode * vp,int flags,enum vgetstate vs)3360 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
3361 {
3362 	int error;
3363 
3364 	if ((flags & LK_INTERLOCK) != 0)
3365 		ASSERT_VI_LOCKED(vp, __func__);
3366 	else
3367 		ASSERT_VI_UNLOCKED(vp, __func__);
3368 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3369 	VNPASS(vp->v_holdcnt > 0, vp);
3370 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3371 
3372 	error = vn_lock(vp, flags);
3373 	if (__predict_false(error != 0)) {
3374 		vget_abort(vp, vs);
3375 		CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3376 		    vp);
3377 		return (error);
3378 	}
3379 
3380 	vget_finish_ref(vp, vs);
3381 	return (0);
3382 }
3383 
3384 void
vget_finish_ref(struct vnode * vp,enum vgetstate vs)3385 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3386 {
3387 	int old;
3388 
3389 	VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3390 	VNPASS(vp->v_holdcnt > 0, vp);
3391 	VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3392 
3393 	if (vs == VGET_USECOUNT)
3394 		return;
3395 
3396 	/*
3397 	 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3398 	 * the vnode around. Otherwise someone else lended their hold count and
3399 	 * we have to drop ours.
3400 	 */
3401 	old = atomic_fetchadd_int(&vp->v_usecount, 1);
3402 	VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3403 	if (old != 0) {
3404 #ifdef INVARIANTS
3405 		old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3406 		VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3407 #else
3408 		refcount_release(&vp->v_holdcnt);
3409 #endif
3410 	}
3411 }
3412 
3413 void
vref(struct vnode * vp)3414 vref(struct vnode *vp)
3415 {
3416 	enum vgetstate vs;
3417 
3418 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3419 	vs = vget_prep(vp);
3420 	vget_finish_ref(vp, vs);
3421 }
3422 
3423 void
vrefact(struct vnode * vp)3424 vrefact(struct vnode *vp)
3425 {
3426 	int old __diagused;
3427 
3428 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3429 	old = refcount_acquire(&vp->v_usecount);
3430 	VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3431 }
3432 
3433 void
vlazy(struct vnode * vp)3434 vlazy(struct vnode *vp)
3435 {
3436 	struct mount *mp;
3437 
3438 	VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3439 
3440 	if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3441 		return;
3442 	/*
3443 	 * We may get here for inactive routines after the vnode got doomed.
3444 	 */
3445 	if (VN_IS_DOOMED(vp))
3446 		return;
3447 	mp = vp->v_mount;
3448 	mtx_lock(&mp->mnt_listmtx);
3449 	if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3450 		vp->v_mflag |= VMP_LAZYLIST;
3451 		TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3452 		mp->mnt_lazyvnodelistsize++;
3453 	}
3454 	mtx_unlock(&mp->mnt_listmtx);
3455 }
3456 
3457 static void
vunlazy(struct vnode * vp)3458 vunlazy(struct vnode *vp)
3459 {
3460 	struct mount *mp;
3461 
3462 	ASSERT_VI_LOCKED(vp, __func__);
3463 	VNPASS(!VN_IS_DOOMED(vp), vp);
3464 
3465 	mp = vp->v_mount;
3466 	mtx_lock(&mp->mnt_listmtx);
3467 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3468 	/*
3469 	 * Don't remove the vnode from the lazy list if another thread
3470 	 * has increased the hold count. It may have re-enqueued the
3471 	 * vnode to the lazy list and is now responsible for its
3472 	 * removal.
3473 	 */
3474 	if (vp->v_holdcnt == 0) {
3475 		vp->v_mflag &= ~VMP_LAZYLIST;
3476 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3477 		mp->mnt_lazyvnodelistsize--;
3478 	}
3479 	mtx_unlock(&mp->mnt_listmtx);
3480 }
3481 
3482 /*
3483  * This routine is only meant to be called from vgonel prior to dooming
3484  * the vnode.
3485  */
3486 static void
vunlazy_gone(struct vnode * vp)3487 vunlazy_gone(struct vnode *vp)
3488 {
3489 	struct mount *mp;
3490 
3491 	ASSERT_VOP_ELOCKED(vp, __func__);
3492 	ASSERT_VI_LOCKED(vp, __func__);
3493 	VNPASS(!VN_IS_DOOMED(vp), vp);
3494 
3495 	if (vp->v_mflag & VMP_LAZYLIST) {
3496 		mp = vp->v_mount;
3497 		mtx_lock(&mp->mnt_listmtx);
3498 		VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3499 		vp->v_mflag &= ~VMP_LAZYLIST;
3500 		TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3501 		mp->mnt_lazyvnodelistsize--;
3502 		mtx_unlock(&mp->mnt_listmtx);
3503 	}
3504 }
3505 
3506 static void
vdefer_inactive(struct vnode * vp)3507 vdefer_inactive(struct vnode *vp)
3508 {
3509 
3510 	ASSERT_VI_LOCKED(vp, __func__);
3511 	VNPASS(vp->v_holdcnt > 0, vp);
3512 	if (VN_IS_DOOMED(vp)) {
3513 		vdropl(vp);
3514 		return;
3515 	}
3516 	if (vp->v_iflag & VI_DEFINACT) {
3517 		VNPASS(vp->v_holdcnt > 1, vp);
3518 		vdropl(vp);
3519 		return;
3520 	}
3521 	if (vp->v_usecount > 0) {
3522 		vp->v_iflag &= ~VI_OWEINACT;
3523 		vdropl(vp);
3524 		return;
3525 	}
3526 	vlazy(vp);
3527 	vp->v_iflag |= VI_DEFINACT;
3528 	VI_UNLOCK(vp);
3529 	atomic_add_long(&deferred_inact, 1);
3530 }
3531 
3532 static void
vdefer_inactive_unlocked(struct vnode * vp)3533 vdefer_inactive_unlocked(struct vnode *vp)
3534 {
3535 
3536 	VI_LOCK(vp);
3537 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
3538 		vdropl(vp);
3539 		return;
3540 	}
3541 	vdefer_inactive(vp);
3542 }
3543 
3544 enum vput_op { VRELE, VPUT, VUNREF };
3545 
3546 /*
3547  * Handle ->v_usecount transitioning to 0.
3548  *
3549  * By releasing the last usecount we take ownership of the hold count which
3550  * provides liveness of the vnode, meaning we have to vdrop.
3551  *
3552  * For all vnodes we may need to perform inactive processing. It requires an
3553  * exclusive lock on the vnode, while it is legal to call here with only a
3554  * shared lock (or no locks). If locking the vnode in an expected manner fails,
3555  * inactive processing gets deferred to the syncer.
3556  *
3557  * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3558  * on the lock being held all the way until VOP_INACTIVE. This in particular
3559  * happens with UFS which adds half-constructed vnodes to the hash, where they
3560  * can be found by other code.
3561  */
3562 static void
vput_final(struct vnode * vp,enum vput_op func)3563 vput_final(struct vnode *vp, enum vput_op func)
3564 {
3565 	int error;
3566 	bool want_unlock;
3567 
3568 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3569 	VNPASS(vp->v_holdcnt > 0, vp);
3570 
3571 	VI_LOCK(vp);
3572 
3573 	/*
3574 	 * By the time we got here someone else might have transitioned
3575 	 * the count back to > 0.
3576 	 */
3577 	if (vp->v_usecount > 0)
3578 		goto out;
3579 
3580 	/*
3581 	 * If the vnode is doomed vgone already performed inactive processing
3582 	 * (if needed).
3583 	 */
3584 	if (VN_IS_DOOMED(vp))
3585 		goto out;
3586 
3587 	if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3588 		goto out;
3589 
3590 	if (vp->v_iflag & VI_DOINGINACT)
3591 		goto out;
3592 
3593 	/*
3594 	 * Locking operations here will drop the interlock and possibly the
3595 	 * vnode lock, opening a window where the vnode can get doomed all the
3596 	 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3597 	 * perform inactive.
3598 	 */
3599 	vp->v_iflag |= VI_OWEINACT;
3600 	want_unlock = false;
3601 	error = 0;
3602 	switch (func) {
3603 	case VRELE:
3604 		switch (VOP_ISLOCKED(vp)) {
3605 		case LK_EXCLUSIVE:
3606 			break;
3607 		case LK_EXCLOTHER:
3608 		case 0:
3609 			want_unlock = true;
3610 			error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3611 			VI_LOCK(vp);
3612 			break;
3613 		default:
3614 			/*
3615 			 * The lock has at least one sharer, but we have no way
3616 			 * to conclude whether this is us. Play it safe and
3617 			 * defer processing.
3618 			 */
3619 			error = EAGAIN;
3620 			break;
3621 		}
3622 		break;
3623 	case VPUT:
3624 		want_unlock = true;
3625 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3626 			error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3627 			    LK_NOWAIT);
3628 			VI_LOCK(vp);
3629 		}
3630 		break;
3631 	case VUNREF:
3632 		if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3633 			error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3634 			VI_LOCK(vp);
3635 		}
3636 		break;
3637 	}
3638 	if (error == 0) {
3639 		if (func == VUNREF) {
3640 			VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3641 			    ("recursive vunref"));
3642 			vp->v_vflag |= VV_UNREF;
3643 		}
3644 		for (;;) {
3645 			error = vinactive(vp);
3646 			if (want_unlock)
3647 				VOP_UNLOCK(vp);
3648 			if (error != ERELOOKUP || !want_unlock)
3649 				break;
3650 			VOP_LOCK(vp, LK_EXCLUSIVE);
3651 		}
3652 		if (func == VUNREF)
3653 			vp->v_vflag &= ~VV_UNREF;
3654 		vdropl(vp);
3655 	} else {
3656 		vdefer_inactive(vp);
3657 	}
3658 	return;
3659 out:
3660 	if (func == VPUT)
3661 		VOP_UNLOCK(vp);
3662 	vdropl(vp);
3663 }
3664 
3665 /*
3666  * Decrement ->v_usecount for a vnode.
3667  *
3668  * Releasing the last use count requires additional processing, see vput_final
3669  * above for details.
3670  *
3671  * Comment above each variant denotes lock state on entry and exit.
3672  */
3673 
3674 /*
3675  * in: any
3676  * out: same as passed in
3677  */
3678 void
vrele(struct vnode * vp)3679 vrele(struct vnode *vp)
3680 {
3681 
3682 	ASSERT_VI_UNLOCKED(vp, __func__);
3683 	if (!refcount_release(&vp->v_usecount))
3684 		return;
3685 	vput_final(vp, VRELE);
3686 }
3687 
3688 /*
3689  * in: locked
3690  * out: unlocked
3691  */
3692 void
vput(struct vnode * vp)3693 vput(struct vnode *vp)
3694 {
3695 
3696 	ASSERT_VOP_LOCKED(vp, __func__);
3697 	ASSERT_VI_UNLOCKED(vp, __func__);
3698 	if (!refcount_release(&vp->v_usecount)) {
3699 		VOP_UNLOCK(vp);
3700 		return;
3701 	}
3702 	vput_final(vp, VPUT);
3703 }
3704 
3705 /*
3706  * in: locked
3707  * out: locked
3708  */
3709 void
vunref(struct vnode * vp)3710 vunref(struct vnode *vp)
3711 {
3712 
3713 	ASSERT_VOP_LOCKED(vp, __func__);
3714 	ASSERT_VI_UNLOCKED(vp, __func__);
3715 	if (!refcount_release(&vp->v_usecount))
3716 		return;
3717 	vput_final(vp, VUNREF);
3718 }
3719 
3720 void
vhold(struct vnode * vp)3721 vhold(struct vnode *vp)
3722 {
3723 	int old;
3724 
3725 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3726 	old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3727 	VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3728 	    ("%s: wrong hold count %d", __func__, old));
3729 	if (old == 0)
3730 		vfs_freevnodes_dec();
3731 }
3732 
3733 void
vholdnz(struct vnode * vp)3734 vholdnz(struct vnode *vp)
3735 {
3736 
3737 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3738 #ifdef INVARIANTS
3739 	int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3740 	VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3741 	    ("%s: wrong hold count %d", __func__, old));
3742 #else
3743 	atomic_add_int(&vp->v_holdcnt, 1);
3744 #endif
3745 }
3746 
3747 /*
3748  * Grab a hold count unless the vnode is freed.
3749  *
3750  * Only use this routine if vfs smr is the only protection you have against
3751  * freeing the vnode.
3752  *
3753  * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3754  * is not set.  After the flag is set the vnode becomes immutable to anyone but
3755  * the thread which managed to set the flag.
3756  *
3757  * It may be tempting to replace the loop with:
3758  * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3759  * if (count & VHOLD_NO_SMR) {
3760  *     backpedal and error out;
3761  * }
3762  *
3763  * However, while this is more performant, it hinders debugging by eliminating
3764  * the previously mentioned invariant.
3765  */
3766 bool
vhold_smr(struct vnode * vp)3767 vhold_smr(struct vnode *vp)
3768 {
3769 	int count;
3770 
3771 	VFS_SMR_ASSERT_ENTERED();
3772 
3773 	count = atomic_load_int(&vp->v_holdcnt);
3774 	for (;;) {
3775 		if (count & VHOLD_NO_SMR) {
3776 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3777 			    ("non-zero hold count with flags %d\n", count));
3778 			return (false);
3779 		}
3780 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3781 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3782 			if (count == 0)
3783 				vfs_freevnodes_dec();
3784 			return (true);
3785 		}
3786 	}
3787 }
3788 
3789 /*
3790  * Hold a free vnode for recycling.
3791  *
3792  * Note: vnode_init references this comment.
3793  *
3794  * Attempts to recycle only need the global vnode list lock and have no use for
3795  * SMR.
3796  *
3797  * However, vnodes get inserted into the global list before they get fully
3798  * initialized and stay there until UMA decides to free the memory. This in
3799  * particular means the target can be found before it becomes usable and after
3800  * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3801  * VHOLD_NO_SMR.
3802  *
3803  * Note: the vnode may gain more references after we transition the count 0->1.
3804  */
3805 static bool
vhold_recycle_free(struct vnode * vp)3806 vhold_recycle_free(struct vnode *vp)
3807 {
3808 	int count;
3809 
3810 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3811 
3812 	count = atomic_load_int(&vp->v_holdcnt);
3813 	for (;;) {
3814 		if (count & VHOLD_NO_SMR) {
3815 			VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3816 			    ("non-zero hold count with flags %d\n", count));
3817 			return (false);
3818 		}
3819 		VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3820 		if (count > 0) {
3821 			return (false);
3822 		}
3823 		if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3824 			vfs_freevnodes_dec();
3825 			return (true);
3826 		}
3827 	}
3828 }
3829 
3830 static void __noinline
vdbatch_process(struct vdbatch * vd)3831 vdbatch_process(struct vdbatch *vd)
3832 {
3833 	struct vnode *vp;
3834 	int i;
3835 
3836 	mtx_assert(&vd->lock, MA_OWNED);
3837 	MPASS(curthread->td_pinned > 0);
3838 	MPASS(vd->index == VDBATCH_SIZE);
3839 
3840 	/*
3841 	 * Attempt to requeue the passed batch, but give up easily.
3842 	 *
3843 	 * Despite batching the mechanism is prone to transient *significant*
3844 	 * lock contention, where vnode_list_mtx becomes the primary bottleneck
3845 	 * if multiple CPUs get here (one real-world example is highly parallel
3846 	 * do-nothing make , which will stat *tons* of vnodes). Since it is
3847 	 * quasi-LRU (read: not that great even if fully honoured) provide an
3848 	 * option to just dodge the problem. Parties which don't like it are
3849 	 * welcome to implement something better.
3850 	 */
3851 	if (vnode_can_skip_requeue) {
3852 		if (!mtx_trylock(&vnode_list_mtx)) {
3853 			counter_u64_add(vnode_skipped_requeues, 1);
3854 			critical_enter();
3855 			for (i = 0; i < VDBATCH_SIZE; i++) {
3856 				vp = vd->tab[i];
3857 				vd->tab[i] = NULL;
3858 				MPASS(vp->v_dbatchcpu != NOCPU);
3859 				vp->v_dbatchcpu = NOCPU;
3860 			}
3861 			vd->index = 0;
3862 			critical_exit();
3863 			return;
3864 
3865 		}
3866 		/* fallthrough to locked processing */
3867 	} else {
3868 		mtx_lock(&vnode_list_mtx);
3869 	}
3870 
3871 	mtx_assert(&vnode_list_mtx, MA_OWNED);
3872 	critical_enter();
3873 	for (i = 0; i < VDBATCH_SIZE; i++) {
3874 		vp = vd->tab[i];
3875 		vd->tab[i] = NULL;
3876 		TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3877 		TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3878 		MPASS(vp->v_dbatchcpu != NOCPU);
3879 		vp->v_dbatchcpu = NOCPU;
3880 	}
3881 	mtx_unlock(&vnode_list_mtx);
3882 	vd->index = 0;
3883 	critical_exit();
3884 }
3885 
3886 static void
vdbatch_enqueue(struct vnode * vp)3887 vdbatch_enqueue(struct vnode *vp)
3888 {
3889 	struct vdbatch *vd;
3890 
3891 	ASSERT_VI_LOCKED(vp, __func__);
3892 	VNPASS(!VN_IS_DOOMED(vp), vp);
3893 
3894 	if (vp->v_dbatchcpu != NOCPU) {
3895 		VI_UNLOCK(vp);
3896 		return;
3897 	}
3898 
3899 	sched_pin();
3900 	vd = DPCPU_PTR(vd);
3901 	mtx_lock(&vd->lock);
3902 	MPASS(vd->index < VDBATCH_SIZE);
3903 	MPASS(vd->tab[vd->index] == NULL);
3904 	/*
3905 	 * A hack: we depend on being pinned so that we know what to put in
3906 	 * ->v_dbatchcpu.
3907 	 */
3908 	vp->v_dbatchcpu = curcpu;
3909 	vd->tab[vd->index] = vp;
3910 	vd->index++;
3911 	VI_UNLOCK(vp);
3912 	if (vd->index == VDBATCH_SIZE)
3913 		vdbatch_process(vd);
3914 	mtx_unlock(&vd->lock);
3915 	sched_unpin();
3916 }
3917 
3918 /*
3919  * This routine must only be called for vnodes which are about to be
3920  * deallocated. Supporting dequeue for arbitrary vndoes would require
3921  * validating that the locked batch matches.
3922  */
3923 static void
vdbatch_dequeue(struct vnode * vp)3924 vdbatch_dequeue(struct vnode *vp)
3925 {
3926 	struct vdbatch *vd;
3927 	int i;
3928 	short cpu;
3929 
3930 	VNPASS(vp->v_type == VBAD || vp->v_type == VNON, vp);
3931 
3932 	cpu = vp->v_dbatchcpu;
3933 	if (cpu == NOCPU)
3934 		return;
3935 
3936 	vd = DPCPU_ID_PTR(cpu, vd);
3937 	mtx_lock(&vd->lock);
3938 	for (i = 0; i < vd->index; i++) {
3939 		if (vd->tab[i] != vp)
3940 			continue;
3941 		vp->v_dbatchcpu = NOCPU;
3942 		vd->index--;
3943 		vd->tab[i] = vd->tab[vd->index];
3944 		vd->tab[vd->index] = NULL;
3945 		break;
3946 	}
3947 	mtx_unlock(&vd->lock);
3948 	/*
3949 	 * Either we dequeued the vnode above or the target CPU beat us to it.
3950 	 */
3951 	MPASS(vp->v_dbatchcpu == NOCPU);
3952 }
3953 
3954 /*
3955  * Drop the hold count of the vnode.
3956  *
3957  * It will only get freed if this is the last hold *and* it has been vgone'd.
3958  *
3959  * Because the vnode vm object keeps a hold reference on the vnode if
3960  * there is at least one resident non-cached page, the vnode cannot
3961  * leave the active list without the page cleanup done.
3962  */
3963 static void __noinline
vdropl_final(struct vnode * vp)3964 vdropl_final(struct vnode *vp)
3965 {
3966 
3967 	ASSERT_VI_LOCKED(vp, __func__);
3968 	VNPASS(VN_IS_DOOMED(vp), vp);
3969 	/*
3970 	 * Set the VHOLD_NO_SMR flag.
3971 	 *
3972 	 * We may be racing against vhold_smr. If they win we can just pretend
3973 	 * we never got this far, they will vdrop later.
3974 	 */
3975 	if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3976 		vfs_freevnodes_inc();
3977 		VI_UNLOCK(vp);
3978 		/*
3979 		 * We lost the aforementioned race. Any subsequent access is
3980 		 * invalid as they might have managed to vdropl on their own.
3981 		 */
3982 		return;
3983 	}
3984 	/*
3985 	 * Don't bump freevnodes as this one is going away.
3986 	 */
3987 	freevnode(vp);
3988 }
3989 
3990 void
vdrop(struct vnode * vp)3991 vdrop(struct vnode *vp)
3992 {
3993 
3994 	ASSERT_VI_UNLOCKED(vp, __func__);
3995 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3996 	if (refcount_release_if_not_last(&vp->v_holdcnt))
3997 		return;
3998 	VI_LOCK(vp);
3999 	vdropl(vp);
4000 }
4001 
4002 static __always_inline void
vdropl_impl(struct vnode * vp,bool enqueue)4003 vdropl_impl(struct vnode *vp, bool enqueue)
4004 {
4005 
4006 	ASSERT_VI_LOCKED(vp, __func__);
4007 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4008 	if (!refcount_release(&vp->v_holdcnt)) {
4009 		VI_UNLOCK(vp);
4010 		return;
4011 	}
4012 	VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
4013 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
4014 	if (VN_IS_DOOMED(vp)) {
4015 		vdropl_final(vp);
4016 		return;
4017 	}
4018 
4019 	vfs_freevnodes_inc();
4020 	if (vp->v_mflag & VMP_LAZYLIST) {
4021 		vunlazy(vp);
4022 	}
4023 
4024 	if (!enqueue) {
4025 		VI_UNLOCK(vp);
4026 		return;
4027 	}
4028 
4029 	/*
4030 	 * Also unlocks the interlock. We can't assert on it as we
4031 	 * released our hold and by now the vnode might have been
4032 	 * freed.
4033 	 */
4034 	vdbatch_enqueue(vp);
4035 }
4036 
4037 void
vdropl(struct vnode * vp)4038 vdropl(struct vnode *vp)
4039 {
4040 
4041 	vdropl_impl(vp, true);
4042 }
4043 
4044 /*
4045  * vdrop a vnode when recycling
4046  *
4047  * This is a special case routine only to be used when recycling, differs from
4048  * regular vdrop by not requeieing the vnode on LRU.
4049  *
4050  * Consider a case where vtryrecycle continuously fails with all vnodes (due to
4051  * e.g., frozen writes on the filesystem), filling the batch and causing it to
4052  * be requeued. Then vnlru will end up revisiting the same vnodes. This is a
4053  * loop which can last for as long as writes are frozen.
4054  */
4055 static void
vdropl_recycle(struct vnode * vp)4056 vdropl_recycle(struct vnode *vp)
4057 {
4058 
4059 	vdropl_impl(vp, false);
4060 }
4061 
4062 static void
vdrop_recycle(struct vnode * vp)4063 vdrop_recycle(struct vnode *vp)
4064 {
4065 
4066 	VI_LOCK(vp);
4067 	vdropl_recycle(vp);
4068 }
4069 
4070 /*
4071  * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
4072  * flags.  DOINGINACT prevents us from recursing in calls to vinactive.
4073  */
4074 static int
vinactivef(struct vnode * vp)4075 vinactivef(struct vnode *vp)
4076 {
4077 	int error;
4078 
4079 	ASSERT_VOP_ELOCKED(vp, "vinactive");
4080 	ASSERT_VI_LOCKED(vp, "vinactive");
4081 	VNPASS((vp->v_iflag & VI_DOINGINACT) == 0, vp);
4082 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4083 	vp->v_iflag |= VI_DOINGINACT;
4084 	vp->v_iflag &= ~VI_OWEINACT;
4085 	VI_UNLOCK(vp);
4086 
4087 	/*
4088 	 * Before moving off the active list, we must be sure that any
4089 	 * modified pages are converted into the vnode's dirty
4090 	 * buffers, since these will no longer be checked once the
4091 	 * vnode is on the inactive list.
4092 	 *
4093 	 * The write-out of the dirty pages is asynchronous.  At the
4094 	 * point that VOP_INACTIVE() is called, there could still be
4095 	 * pending I/O and dirty pages in the object.
4096 	 */
4097 	if ((vp->v_vflag & VV_NOSYNC) == 0)
4098 		vnode_pager_clean_async(vp);
4099 
4100 	error = VOP_INACTIVE(vp);
4101 	VI_LOCK(vp);
4102 	VNPASS(vp->v_iflag & VI_DOINGINACT, vp);
4103 	vp->v_iflag &= ~VI_DOINGINACT;
4104 	return (error);
4105 }
4106 
4107 int
vinactive(struct vnode * vp)4108 vinactive(struct vnode *vp)
4109 {
4110 
4111 	ASSERT_VOP_ELOCKED(vp, "vinactive");
4112 	ASSERT_VI_LOCKED(vp, "vinactive");
4113 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4114 
4115 	if ((vp->v_iflag & VI_OWEINACT) == 0)
4116 		return (0);
4117 	if (vp->v_iflag & VI_DOINGINACT)
4118 		return (0);
4119 	if (vp->v_usecount > 0) {
4120 		vp->v_iflag &= ~VI_OWEINACT;
4121 		return (0);
4122 	}
4123 	return (vinactivef(vp));
4124 }
4125 
4126 /*
4127  * Remove any vnodes in the vnode table belonging to mount point mp.
4128  *
4129  * If FORCECLOSE is not specified, there should not be any active ones,
4130  * return error if any are found (nb: this is a user error, not a
4131  * system error). If FORCECLOSE is specified, detach any active vnodes
4132  * that are found.
4133  *
4134  * If WRITECLOSE is set, only flush out regular file vnodes open for
4135  * writing.
4136  *
4137  * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
4138  *
4139  * `rootrefs' specifies the base reference count for the root vnode
4140  * of this filesystem. The root vnode is considered busy if its
4141  * v_usecount exceeds this value. On a successful return, vflush(, td)
4142  * will call vrele() on the root vnode exactly rootrefs times.
4143  * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
4144  * be zero.
4145  */
4146 #ifdef DIAGNOSTIC
4147 static int busyprt = 0;		/* print out busy vnodes */
4148 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
4149 #endif
4150 
4151 int
vflush(struct mount * mp,int rootrefs,int flags,struct thread * td)4152 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
4153 {
4154 	struct vnode *vp, *mvp, *rootvp = NULL;
4155 	struct vattr vattr;
4156 	int busy = 0, error;
4157 
4158 	CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
4159 	    rootrefs, flags);
4160 	if (rootrefs > 0) {
4161 		KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
4162 		    ("vflush: bad args"));
4163 		/*
4164 		 * Get the filesystem root vnode. We can vput() it
4165 		 * immediately, since with rootrefs > 0, it won't go away.
4166 		 */
4167 		if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
4168 			CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
4169 			    __func__, error);
4170 			return (error);
4171 		}
4172 		vput(rootvp);
4173 	}
4174 loop:
4175 	MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
4176 		vholdl(vp);
4177 		error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
4178 		if (error) {
4179 			vdrop(vp);
4180 			MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
4181 			goto loop;
4182 		}
4183 		/*
4184 		 * Skip over a vnodes marked VV_SYSTEM.
4185 		 */
4186 		if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
4187 			VOP_UNLOCK(vp);
4188 			vdrop(vp);
4189 			continue;
4190 		}
4191 		/*
4192 		 * If WRITECLOSE is set, flush out unlinked but still open
4193 		 * files (even if open only for reading) and regular file
4194 		 * vnodes open for writing.
4195 		 */
4196 		if (flags & WRITECLOSE) {
4197 			vnode_pager_clean_async(vp);
4198 			do {
4199 				error = VOP_FSYNC(vp, MNT_WAIT, td);
4200 			} while (error == ERELOOKUP);
4201 			if (error != 0) {
4202 				VOP_UNLOCK(vp);
4203 				vdrop(vp);
4204 				MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
4205 				return (error);
4206 			}
4207 			error = VOP_GETATTR(vp, &vattr, td->td_ucred);
4208 			VI_LOCK(vp);
4209 
4210 			if ((vp->v_type == VNON ||
4211 			    (error == 0 && vattr.va_nlink > 0)) &&
4212 			    (vp->v_writecount <= 0 || vp->v_type != VREG)) {
4213 				VOP_UNLOCK(vp);
4214 				vdropl(vp);
4215 				continue;
4216 			}
4217 		} else
4218 			VI_LOCK(vp);
4219 		/*
4220 		 * With v_usecount == 0, all we need to do is clear out the
4221 		 * vnode data structures and we are done.
4222 		 *
4223 		 * If FORCECLOSE is set, forcibly close the vnode.
4224 		 */
4225 		if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
4226 			vgonel(vp);
4227 		} else {
4228 			busy++;
4229 #ifdef DIAGNOSTIC
4230 			if (busyprt)
4231 				vn_printf(vp, "vflush: busy vnode ");
4232 #endif
4233 		}
4234 		VOP_UNLOCK(vp);
4235 		vdropl(vp);
4236 	}
4237 	if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
4238 		/*
4239 		 * If just the root vnode is busy, and if its refcount
4240 		 * is equal to `rootrefs', then go ahead and kill it.
4241 		 */
4242 		VI_LOCK(rootvp);
4243 		KASSERT(busy > 0, ("vflush: not busy"));
4244 		VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
4245 		    ("vflush: usecount %d < rootrefs %d",
4246 		     rootvp->v_usecount, rootrefs));
4247 		if (busy == 1 && rootvp->v_usecount == rootrefs) {
4248 			VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
4249 			vgone(rootvp);
4250 			VOP_UNLOCK(rootvp);
4251 			busy = 0;
4252 		} else
4253 			VI_UNLOCK(rootvp);
4254 	}
4255 	if (busy) {
4256 		CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
4257 		    busy);
4258 		return (EBUSY);
4259 	}
4260 	for (; rootrefs > 0; rootrefs--)
4261 		vrele(rootvp);
4262 	return (0);
4263 }
4264 
4265 /*
4266  * Recycle an unused vnode.
4267  */
4268 int
vrecycle(struct vnode * vp)4269 vrecycle(struct vnode *vp)
4270 {
4271 	int recycled;
4272 
4273 	VI_LOCK(vp);
4274 	recycled = vrecyclel(vp);
4275 	VI_UNLOCK(vp);
4276 	return (recycled);
4277 }
4278 
4279 /*
4280  * vrecycle, with the vp interlock held.
4281  */
4282 int
vrecyclel(struct vnode * vp)4283 vrecyclel(struct vnode *vp)
4284 {
4285 	int recycled;
4286 
4287 	ASSERT_VOP_ELOCKED(vp, __func__);
4288 	ASSERT_VI_LOCKED(vp, __func__);
4289 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4290 	recycled = 0;
4291 	if (vp->v_usecount == 0) {
4292 		recycled = 1;
4293 		vgonel(vp);
4294 	}
4295 	return (recycled);
4296 }
4297 
4298 /*
4299  * Eliminate all activity associated with a vnode
4300  * in preparation for reuse.
4301  */
4302 void
vgone(struct vnode * vp)4303 vgone(struct vnode *vp)
4304 {
4305 	VI_LOCK(vp);
4306 	vgonel(vp);
4307 	VI_UNLOCK(vp);
4308 }
4309 
4310 /*
4311  * Notify upper mounts about reclaimed or unlinked vnode.
4312  */
4313 void
vfs_notify_upper(struct vnode * vp,enum vfs_notify_upper_type event)4314 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event)
4315 {
4316 	struct mount *mp;
4317 	struct mount_upper_node *ump;
4318 
4319 	mp = atomic_load_ptr(&vp->v_mount);
4320 	if (mp == NULL)
4321 		return;
4322 	if (TAILQ_EMPTY(&mp->mnt_notify))
4323 		return;
4324 
4325 	MNT_ILOCK(mp);
4326 	mp->mnt_upper_pending++;
4327 	KASSERT(mp->mnt_upper_pending > 0,
4328 	    ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending));
4329 	TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) {
4330 		MNT_IUNLOCK(mp);
4331 		switch (event) {
4332 		case VFS_NOTIFY_UPPER_RECLAIM:
4333 			VFS_RECLAIM_LOWERVP(ump->mp, vp);
4334 			break;
4335 		case VFS_NOTIFY_UPPER_UNLINK:
4336 			VFS_UNLINK_LOWERVP(ump->mp, vp);
4337 			break;
4338 		}
4339 		MNT_ILOCK(mp);
4340 	}
4341 	mp->mnt_upper_pending--;
4342 	if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 &&
4343 	    mp->mnt_upper_pending == 0) {
4344 		mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER;
4345 		wakeup(&mp->mnt_uppers);
4346 	}
4347 	MNT_IUNLOCK(mp);
4348 }
4349 
4350 /*
4351  * vgone, with the vp interlock held.
4352  */
4353 static void
vgonel(struct vnode * vp)4354 vgonel(struct vnode *vp)
4355 {
4356 	struct thread *td;
4357 	struct mount *mp;
4358 	vm_object_t object;
4359 	bool active, doinginact, oweinact;
4360 
4361 	ASSERT_VOP_ELOCKED(vp, "vgonel");
4362 	ASSERT_VI_LOCKED(vp, "vgonel");
4363 	VNASSERT(vp->v_holdcnt, vp,
4364 	    ("vgonel: vp %p has no reference.", vp));
4365 	CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4366 	td = curthread;
4367 
4368 	/*
4369 	 * Don't vgonel if we're already doomed.
4370 	 */
4371 	if (VN_IS_DOOMED(vp)) {
4372 		VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \
4373 		    vn_get_state(vp) == VSTATE_DEAD, vp);
4374 		return;
4375 	}
4376 	/*
4377 	 * Paired with freevnode.
4378 	 */
4379 	vn_seqc_write_begin_locked(vp);
4380 	vunlazy_gone(vp);
4381 	vn_irflag_set_locked(vp, VIRF_DOOMED);
4382 	vn_set_state(vp, VSTATE_DESTROYING);
4383 
4384 	/*
4385 	 * Check to see if the vnode is in use.  If so, we have to
4386 	 * call VOP_CLOSE() and VOP_INACTIVE().
4387 	 *
4388 	 * It could be that VOP_INACTIVE() requested reclamation, in
4389 	 * which case we should avoid recursion, so check
4390 	 * VI_DOINGINACT.  This is not precise but good enough.
4391 	 */
4392 	active = vp->v_usecount > 0;
4393 	oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4394 	doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
4395 
4396 	/*
4397 	 * If we need to do inactive VI_OWEINACT will be set.
4398 	 */
4399 	if (vp->v_iflag & VI_DEFINACT) {
4400 		VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
4401 		vp->v_iflag &= ~VI_DEFINACT;
4402 		vdropl(vp);
4403 	} else {
4404 		VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
4405 		VI_UNLOCK(vp);
4406 	}
4407 	cache_purge_vgone(vp);
4408 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
4409 
4410 	/*
4411 	 * If purging an active vnode, it must be closed and
4412 	 * deactivated before being reclaimed.
4413 	 */
4414 	if (active)
4415 		VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
4416 	if (!doinginact) {
4417 		do {
4418 			if (oweinact || active) {
4419 				VI_LOCK(vp);
4420 				vinactivef(vp);
4421 				oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4422 				VI_UNLOCK(vp);
4423 			}
4424 		} while (oweinact);
4425 	}
4426 	if (vp->v_type == VSOCK)
4427 		vfs_unp_reclaim(vp);
4428 
4429 	/*
4430 	 * Clean out any buffers associated with the vnode.
4431 	 * If the flush fails, just toss the buffers.
4432 	 */
4433 	mp = NULL;
4434 	if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4435 		(void) vn_start_secondary_write(vp, &mp, V_WAIT);
4436 	if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4437 		while (vinvalbuf(vp, 0, 0, 0) != 0)
4438 			;
4439 	}
4440 
4441 	BO_LOCK(&vp->v_bufobj);
4442 	KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4443 	    vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4444 	    TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4445 	    vp->v_bufobj.bo_clean.bv_cnt == 0,
4446 	    ("vp %p bufobj not invalidated", vp));
4447 
4448 	/*
4449 	 * For VMIO bufobj, BO_DEAD is set later, or in
4450 	 * vm_object_terminate() after the object's page queue is
4451 	 * flushed.
4452 	 */
4453 	object = vp->v_bufobj.bo_object;
4454 	if (object == NULL)
4455 		vp->v_bufobj.bo_flag |= BO_DEAD;
4456 	BO_UNLOCK(&vp->v_bufobj);
4457 
4458 	/*
4459 	 * Handle the VM part.  Tmpfs handles v_object on its own (the
4460 	 * OBJT_VNODE check).  Nullfs or other bypassing filesystems
4461 	 * should not touch the object borrowed from the lower vnode
4462 	 * (the handle check).
4463 	 */
4464 	if (object != NULL && object->type == OBJT_VNODE &&
4465 	    object->handle == vp)
4466 		vnode_destroy_vobject(vp);
4467 
4468 	/*
4469 	 * Reclaim the vnode.
4470 	 */
4471 	if (VOP_RECLAIM(vp))
4472 		panic("vgone: cannot reclaim");
4473 	if (mp != NULL)
4474 		vn_finished_secondary_write(mp);
4475 	VNASSERT(vp->v_object == NULL, vp,
4476 	    ("vop_reclaim left v_object vp=%p", vp));
4477 	/*
4478 	 * Clear the advisory locks and wake up waiting threads.
4479 	 */
4480 	if (vp->v_lockf != NULL) {
4481 		(void)VOP_ADVLOCKPURGE(vp);
4482 		vp->v_lockf = NULL;
4483 	}
4484 	/*
4485 	 * Delete from old mount point vnode list.
4486 	 */
4487 	if (vp->v_mount == NULL) {
4488 		VI_LOCK(vp);
4489 	} else {
4490 		delmntque(vp);
4491 		ASSERT_VI_LOCKED(vp, "vgonel 2");
4492 	}
4493 	/*
4494 	 * Done with purge, reset to the standard lock and invalidate
4495 	 * the vnode.
4496 	 */
4497 	vp->v_vnlock = &vp->v_lock;
4498 	vp->v_op = &dead_vnodeops;
4499 	vp->v_type = VBAD;
4500 	vn_set_state(vp, VSTATE_DEAD);
4501 }
4502 
4503 /*
4504  * Print out a description of a vnode.
4505  */
4506 static const char *const vtypename[] = {
4507 	[VNON] = "VNON",
4508 	[VREG] = "VREG",
4509 	[VDIR] = "VDIR",
4510 	[VBLK] = "VBLK",
4511 	[VCHR] = "VCHR",
4512 	[VLNK] = "VLNK",
4513 	[VSOCK] = "VSOCK",
4514 	[VFIFO] = "VFIFO",
4515 	[VBAD] = "VBAD",
4516 	[VMARKER] = "VMARKER",
4517 };
4518 _Static_assert(nitems(vtypename) == VLASTTYPE + 1,
4519     "vnode type name not added to vtypename");
4520 
4521 static const char *const vstatename[] = {
4522 	[VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED",
4523 	[VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED",
4524 	[VSTATE_DESTROYING] = "VSTATE_DESTROYING",
4525 	[VSTATE_DEAD] = "VSTATE_DEAD",
4526 };
4527 _Static_assert(nitems(vstatename) == VLASTSTATE + 1,
4528     "vnode state name not added to vstatename");
4529 
4530 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4531     "new hold count flag not added to vn_printf");
4532 
4533 void
vn_printf(struct vnode * vp,const char * fmt,...)4534 vn_printf(struct vnode *vp, const char *fmt, ...)
4535 {
4536 	va_list ap;
4537 	char buf[256], buf2[16];
4538 	u_long flags;
4539 	u_int holdcnt;
4540 	short irflag;
4541 
4542 	va_start(ap, fmt);
4543 	vprintf(fmt, ap);
4544 	va_end(ap);
4545 	printf("%p: ", (void *)vp);
4546 	printf("type %s state %s op %p\n", vtypename[vp->v_type],
4547 	    vstatename[vp->v_state], vp->v_op);
4548 	holdcnt = atomic_load_int(&vp->v_holdcnt);
4549 	printf("    usecount %d, writecount %d, refcount %d seqc users %d",
4550 	    vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4551 	    vp->v_seqc_users);
4552 	switch (vp->v_type) {
4553 	case VDIR:
4554 		printf(" mountedhere %p\n", vp->v_mountedhere);
4555 		break;
4556 	case VCHR:
4557 		printf(" rdev %p\n", vp->v_rdev);
4558 		break;
4559 	case VSOCK:
4560 		printf(" socket %p\n", vp->v_unpcb);
4561 		break;
4562 	case VFIFO:
4563 		printf(" fifoinfo %p\n", vp->v_fifoinfo);
4564 		break;
4565 	default:
4566 		printf("\n");
4567 		break;
4568 	}
4569 	buf[0] = '\0';
4570 	buf[1] = '\0';
4571 	if (holdcnt & VHOLD_NO_SMR)
4572 		strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4573 	printf("    hold count flags (%s)\n", buf + 1);
4574 
4575 	buf[0] = '\0';
4576 	buf[1] = '\0';
4577 	irflag = vn_irflag_read(vp);
4578 	if (irflag & VIRF_DOOMED)
4579 		strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4580 	if (irflag & VIRF_PGREAD)
4581 		strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4582 	if (irflag & VIRF_MOUNTPOINT)
4583 		strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4584 	if (irflag & VIRF_TEXT_REF)
4585 		strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf));
4586 	flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF);
4587 	if (flags != 0) {
4588 		snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4589 		strlcat(buf, buf2, sizeof(buf));
4590 	}
4591 	if (vp->v_vflag & VV_ROOT)
4592 		strlcat(buf, "|VV_ROOT", sizeof(buf));
4593 	if (vp->v_vflag & VV_ISTTY)
4594 		strlcat(buf, "|VV_ISTTY", sizeof(buf));
4595 	if (vp->v_vflag & VV_NOSYNC)
4596 		strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4597 	if (vp->v_vflag & VV_ETERNALDEV)
4598 		strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4599 	if (vp->v_vflag & VV_CACHEDLABEL)
4600 		strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4601 	if (vp->v_vflag & VV_VMSIZEVNLOCK)
4602 		strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4603 	if (vp->v_vflag & VV_COPYONWRITE)
4604 		strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4605 	if (vp->v_vflag & VV_SYSTEM)
4606 		strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4607 	if (vp->v_vflag & VV_PROCDEP)
4608 		strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4609 	if (vp->v_vflag & VV_DELETED)
4610 		strlcat(buf, "|VV_DELETED", sizeof(buf));
4611 	if (vp->v_vflag & VV_MD)
4612 		strlcat(buf, "|VV_MD", sizeof(buf));
4613 	if (vp->v_vflag & VV_FORCEINSMQ)
4614 		strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4615 	if (vp->v_vflag & VV_READLINK)
4616 		strlcat(buf, "|VV_READLINK", sizeof(buf));
4617 	flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4618 	    VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4619 	    VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK);
4620 	if (flags != 0) {
4621 		snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4622 		strlcat(buf, buf2, sizeof(buf));
4623 	}
4624 	if (vp->v_iflag & VI_MOUNT)
4625 		strlcat(buf, "|VI_MOUNT", sizeof(buf));
4626 	if (vp->v_iflag & VI_DOINGINACT)
4627 		strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4628 	if (vp->v_iflag & VI_OWEINACT)
4629 		strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4630 	if (vp->v_iflag & VI_DEFINACT)
4631 		strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4632 	if (vp->v_iflag & VI_FOPENING)
4633 		strlcat(buf, "|VI_FOPENING", sizeof(buf));
4634 	flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT |
4635 	    VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4636 	if (flags != 0) {
4637 		snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4638 		strlcat(buf, buf2, sizeof(buf));
4639 	}
4640 	if (vp->v_mflag & VMP_LAZYLIST)
4641 		strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4642 	flags = vp->v_mflag & ~(VMP_LAZYLIST);
4643 	if (flags != 0) {
4644 		snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4645 		strlcat(buf, buf2, sizeof(buf));
4646 	}
4647 	printf("    flags (%s)", buf + 1);
4648 	if (mtx_owned(VI_MTX(vp)))
4649 		printf(" VI_LOCKed");
4650 	printf("\n");
4651 	if (vp->v_object != NULL)
4652 		printf("    v_object %p ref %d pages %d "
4653 		    "cleanbuf %d dirtybuf %d\n",
4654 		    vp->v_object, vp->v_object->ref_count,
4655 		    vp->v_object->resident_page_count,
4656 		    vp->v_bufobj.bo_clean.bv_cnt,
4657 		    vp->v_bufobj.bo_dirty.bv_cnt);
4658 	printf("    ");
4659 	lockmgr_printinfo(vp->v_vnlock);
4660 	if (vp->v_data != NULL)
4661 		VOP_PRINT(vp);
4662 }
4663 
4664 #ifdef DDB
4665 /*
4666  * List all of the locked vnodes in the system.
4667  * Called when debugging the kernel.
4668  */
DB_SHOW_COMMAND_FLAGS(lockedvnods,lockedvnodes,DB_CMD_MEMSAFE)4669 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE)
4670 {
4671 	struct mount *mp;
4672 	struct vnode *vp;
4673 
4674 	/*
4675 	 * Note: because this is DDB, we can't obey the locking semantics
4676 	 * for these structures, which means we could catch an inconsistent
4677 	 * state and dereference a nasty pointer.  Not much to be done
4678 	 * about that.
4679 	 */
4680 	db_printf("Locked vnodes\n");
4681 	TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4682 		TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4683 			if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4684 				vn_printf(vp, "vnode ");
4685 		}
4686 	}
4687 }
4688 
4689 /*
4690  * Show details about the given vnode.
4691  */
DB_SHOW_COMMAND(vnode,db_show_vnode)4692 DB_SHOW_COMMAND(vnode, db_show_vnode)
4693 {
4694 	struct vnode *vp;
4695 
4696 	if (!have_addr)
4697 		return;
4698 	vp = (struct vnode *)addr;
4699 	vn_printf(vp, "vnode ");
4700 }
4701 
4702 /*
4703  * Show details about the given mount point.
4704  */
DB_SHOW_COMMAND(mount,db_show_mount)4705 DB_SHOW_COMMAND(mount, db_show_mount)
4706 {
4707 	struct mount *mp;
4708 	struct vfsopt *opt;
4709 	struct statfs *sp;
4710 	struct vnode *vp;
4711 	char buf[512];
4712 	uint64_t mflags;
4713 	u_int flags;
4714 
4715 	if (!have_addr) {
4716 		/* No address given, print short info about all mount points. */
4717 		TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4718 			db_printf("%p %s on %s (%s)\n", mp,
4719 			    mp->mnt_stat.f_mntfromname,
4720 			    mp->mnt_stat.f_mntonname,
4721 			    mp->mnt_stat.f_fstypename);
4722 			if (db_pager_quit)
4723 				break;
4724 		}
4725 		db_printf("\nMore info: show mount <addr>\n");
4726 		return;
4727 	}
4728 
4729 	mp = (struct mount *)addr;
4730 	db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4731 	    mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4732 
4733 	buf[0] = '\0';
4734 	mflags = mp->mnt_flag;
4735 #define	MNT_FLAG(flag)	do {						\
4736 	if (mflags & (flag)) {						\
4737 		if (buf[0] != '\0')					\
4738 			strlcat(buf, ", ", sizeof(buf));		\
4739 		strlcat(buf, (#flag) + 4, sizeof(buf));			\
4740 		mflags &= ~(flag);					\
4741 	}								\
4742 } while (0)
4743 	MNT_FLAG(MNT_RDONLY);
4744 	MNT_FLAG(MNT_SYNCHRONOUS);
4745 	MNT_FLAG(MNT_NOEXEC);
4746 	MNT_FLAG(MNT_NOSUID);
4747 	MNT_FLAG(MNT_NFS4ACLS);
4748 	MNT_FLAG(MNT_UNION);
4749 	MNT_FLAG(MNT_ASYNC);
4750 	MNT_FLAG(MNT_SUIDDIR);
4751 	MNT_FLAG(MNT_SOFTDEP);
4752 	MNT_FLAG(MNT_NOSYMFOLLOW);
4753 	MNT_FLAG(MNT_GJOURNAL);
4754 	MNT_FLAG(MNT_MULTILABEL);
4755 	MNT_FLAG(MNT_ACLS);
4756 	MNT_FLAG(MNT_NOATIME);
4757 	MNT_FLAG(MNT_NOCLUSTERR);
4758 	MNT_FLAG(MNT_NOCLUSTERW);
4759 	MNT_FLAG(MNT_SUJ);
4760 	MNT_FLAG(MNT_EXRDONLY);
4761 	MNT_FLAG(MNT_EXPORTED);
4762 	MNT_FLAG(MNT_DEFEXPORTED);
4763 	MNT_FLAG(MNT_EXPORTANON);
4764 	MNT_FLAG(MNT_EXKERB);
4765 	MNT_FLAG(MNT_EXPUBLIC);
4766 	MNT_FLAG(MNT_LOCAL);
4767 	MNT_FLAG(MNT_QUOTA);
4768 	MNT_FLAG(MNT_ROOTFS);
4769 	MNT_FLAG(MNT_USER);
4770 	MNT_FLAG(MNT_IGNORE);
4771 	MNT_FLAG(MNT_UPDATE);
4772 	MNT_FLAG(MNT_DELEXPORT);
4773 	MNT_FLAG(MNT_RELOAD);
4774 	MNT_FLAG(MNT_FORCE);
4775 	MNT_FLAG(MNT_SNAPSHOT);
4776 	MNT_FLAG(MNT_BYFSID);
4777 #undef MNT_FLAG
4778 	if (mflags != 0) {
4779 		if (buf[0] != '\0')
4780 			strlcat(buf, ", ", sizeof(buf));
4781 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4782 		    "0x%016jx", mflags);
4783 	}
4784 	db_printf("    mnt_flag = %s\n", buf);
4785 
4786 	buf[0] = '\0';
4787 	flags = mp->mnt_kern_flag;
4788 #define	MNT_KERN_FLAG(flag)	do {					\
4789 	if (flags & (flag)) {						\
4790 		if (buf[0] != '\0')					\
4791 			strlcat(buf, ", ", sizeof(buf));		\
4792 		strlcat(buf, (#flag) + 5, sizeof(buf));			\
4793 		flags &= ~(flag);					\
4794 	}								\
4795 } while (0)
4796 	MNT_KERN_FLAG(MNTK_UNMOUNTF);
4797 	MNT_KERN_FLAG(MNTK_ASYNC);
4798 	MNT_KERN_FLAG(MNTK_SOFTDEP);
4799 	MNT_KERN_FLAG(MNTK_NOMSYNC);
4800 	MNT_KERN_FLAG(MNTK_DRAINING);
4801 	MNT_KERN_FLAG(MNTK_REFEXPIRE);
4802 	MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4803 	MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4804 	MNT_KERN_FLAG(MNTK_NO_IOPF);
4805 	MNT_KERN_FLAG(MNTK_RECURSE);
4806 	MNT_KERN_FLAG(MNTK_UPPER_WAITER);
4807 	MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE);
4808 	MNT_KERN_FLAG(MNTK_USES_BCACHE);
4809 	MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG);
4810 	MNT_KERN_FLAG(MNTK_FPLOOKUP);
4811 	MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER);
4812 	MNT_KERN_FLAG(MNTK_NOASYNC);
4813 	MNT_KERN_FLAG(MNTK_UNMOUNT);
4814 	MNT_KERN_FLAG(MNTK_MWAIT);
4815 	MNT_KERN_FLAG(MNTK_SUSPEND);
4816 	MNT_KERN_FLAG(MNTK_SUSPEND2);
4817 	MNT_KERN_FLAG(MNTK_SUSPENDED);
4818 	MNT_KERN_FLAG(MNTK_NULL_NOCACHE);
4819 	MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4820 #undef MNT_KERN_FLAG
4821 	if (flags != 0) {
4822 		if (buf[0] != '\0')
4823 			strlcat(buf, ", ", sizeof(buf));
4824 		snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4825 		    "0x%08x", flags);
4826 	}
4827 	db_printf("    mnt_kern_flag = %s\n", buf);
4828 
4829 	db_printf("    mnt_opt = ");
4830 	opt = TAILQ_FIRST(mp->mnt_opt);
4831 	if (opt != NULL) {
4832 		db_printf("%s", opt->name);
4833 		opt = TAILQ_NEXT(opt, link);
4834 		while (opt != NULL) {
4835 			db_printf(", %s", opt->name);
4836 			opt = TAILQ_NEXT(opt, link);
4837 		}
4838 	}
4839 	db_printf("\n");
4840 
4841 	sp = &mp->mnt_stat;
4842 	db_printf("    mnt_stat = { version=%u type=%u flags=0x%016jx "
4843 	    "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4844 	    "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4845 	    "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4846 	    (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4847 	    (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4848 	    (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4849 	    (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4850 	    (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4851 	    (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4852 	    (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4853 	    (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4854 
4855 	db_printf("    mnt_cred = { uid=%u ruid=%u",
4856 	    (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4857 	if (jailed(mp->mnt_cred))
4858 		db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4859 	db_printf(" }\n");
4860 	db_printf("    mnt_ref = %d (with %d in the struct)\n",
4861 	    vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4862 	db_printf("    mnt_gen = %d\n", mp->mnt_gen);
4863 	db_printf("    mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4864 	db_printf("    mnt_lazyvnodelistsize = %d\n",
4865 	    mp->mnt_lazyvnodelistsize);
4866 	db_printf("    mnt_writeopcount = %d (with %d in the struct)\n",
4867 	    vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4868 	db_printf("    mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4869 	db_printf("    mnt_hashseed = %u\n", mp->mnt_hashseed);
4870 	db_printf("    mnt_lockref = %d (with %d in the struct)\n",
4871 	    vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4872 	db_printf("    mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4873 	db_printf("    mnt_secondary_accwrites = %d\n",
4874 	    mp->mnt_secondary_accwrites);
4875 	db_printf("    mnt_gjprovider = %s\n",
4876 	    mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4877 	db_printf("    mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4878 
4879 	db_printf("\n\nList of active vnodes\n");
4880 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4881 		if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4882 			vn_printf(vp, "vnode ");
4883 			if (db_pager_quit)
4884 				break;
4885 		}
4886 	}
4887 	db_printf("\n\nList of inactive vnodes\n");
4888 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4889 		if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4890 			vn_printf(vp, "vnode ");
4891 			if (db_pager_quit)
4892 				break;
4893 		}
4894 	}
4895 }
4896 #endif	/* DDB */
4897 
4898 /*
4899  * Fill in a struct xvfsconf based on a struct vfsconf.
4900  */
4901 static int
vfsconf2x(struct sysctl_req * req,struct vfsconf * vfsp)4902 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4903 {
4904 	struct xvfsconf xvfsp;
4905 
4906 	bzero(&xvfsp, sizeof(xvfsp));
4907 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4908 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4909 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4910 	xvfsp.vfc_flags = vfsp->vfc_flags;
4911 	/*
4912 	 * These are unused in userland, we keep them
4913 	 * to not break binary compatibility.
4914 	 */
4915 	xvfsp.vfc_vfsops = NULL;
4916 	xvfsp.vfc_next = NULL;
4917 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4918 }
4919 
4920 #ifdef COMPAT_FREEBSD32
4921 struct xvfsconf32 {
4922 	uint32_t	vfc_vfsops;
4923 	char		vfc_name[MFSNAMELEN];
4924 	int32_t		vfc_typenum;
4925 	int32_t		vfc_refcount;
4926 	int32_t		vfc_flags;
4927 	uint32_t	vfc_next;
4928 };
4929 
4930 static int
vfsconf2x32(struct sysctl_req * req,struct vfsconf * vfsp)4931 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4932 {
4933 	struct xvfsconf32 xvfsp;
4934 
4935 	bzero(&xvfsp, sizeof(xvfsp));
4936 	strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4937 	xvfsp.vfc_typenum = vfsp->vfc_typenum;
4938 	xvfsp.vfc_refcount = vfsp->vfc_refcount;
4939 	xvfsp.vfc_flags = vfsp->vfc_flags;
4940 	return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4941 }
4942 #endif
4943 
4944 /*
4945  * Top level filesystem related information gathering.
4946  */
4947 static int
sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)4948 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4949 {
4950 	struct vfsconf *vfsp;
4951 	int error;
4952 
4953 	error = 0;
4954 	vfsconf_slock();
4955 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4956 #ifdef COMPAT_FREEBSD32
4957 		if (req->flags & SCTL_MASK32)
4958 			error = vfsconf2x32(req, vfsp);
4959 		else
4960 #endif
4961 			error = vfsconf2x(req, vfsp);
4962 		if (error)
4963 			break;
4964 	}
4965 	vfsconf_sunlock();
4966 	return (error);
4967 }
4968 
4969 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4970     CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4971     "S,xvfsconf", "List of all configured filesystems");
4972 
4973 #ifndef BURN_BRIDGES
4974 static int	sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4975 
4976 static int
vfs_sysctl(SYSCTL_HANDLER_ARGS)4977 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4978 {
4979 	int *name = (int *)arg1 - 1;	/* XXX */
4980 	u_int namelen = arg2 + 1;	/* XXX */
4981 	struct vfsconf *vfsp;
4982 
4983 	log(LOG_WARNING, "userland calling deprecated sysctl, "
4984 	    "please rebuild world\n");
4985 
4986 #if 1 || defined(COMPAT_PRELITE2)
4987 	/* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4988 	if (namelen == 1)
4989 		return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4990 #endif
4991 
4992 	switch (name[1]) {
4993 	case VFS_MAXTYPENUM:
4994 		if (namelen != 2)
4995 			return (ENOTDIR);
4996 		return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4997 	case VFS_CONF:
4998 		if (namelen != 3)
4999 			return (ENOTDIR);	/* overloaded */
5000 		vfsconf_slock();
5001 		TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
5002 			if (vfsp->vfc_typenum == name[2])
5003 				break;
5004 		}
5005 		vfsconf_sunlock();
5006 		if (vfsp == NULL)
5007 			return (EOPNOTSUPP);
5008 #ifdef COMPAT_FREEBSD32
5009 		if (req->flags & SCTL_MASK32)
5010 			return (vfsconf2x32(req, vfsp));
5011 		else
5012 #endif
5013 			return (vfsconf2x(req, vfsp));
5014 	}
5015 	return (EOPNOTSUPP);
5016 }
5017 
5018 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
5019     CTLFLAG_MPSAFE, vfs_sysctl,
5020     "Generic filesystem");
5021 
5022 #if 1 || defined(COMPAT_PRELITE2)
5023 
5024 static int
sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)5025 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
5026 {
5027 	int error;
5028 	struct vfsconf *vfsp;
5029 	struct ovfsconf ovfs;
5030 
5031 	vfsconf_slock();
5032 	TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
5033 		bzero(&ovfs, sizeof(ovfs));
5034 		ovfs.vfc_vfsops = vfsp->vfc_vfsops;	/* XXX used as flag */
5035 		strcpy(ovfs.vfc_name, vfsp->vfc_name);
5036 		ovfs.vfc_index = vfsp->vfc_typenum;
5037 		ovfs.vfc_refcount = vfsp->vfc_refcount;
5038 		ovfs.vfc_flags = vfsp->vfc_flags;
5039 		error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
5040 		if (error != 0) {
5041 			vfsconf_sunlock();
5042 			return (error);
5043 		}
5044 	}
5045 	vfsconf_sunlock();
5046 	return (0);
5047 }
5048 
5049 #endif /* 1 || COMPAT_PRELITE2 */
5050 #endif /* !BURN_BRIDGES */
5051 
5052 static void
unmount_or_warn(struct mount * mp)5053 unmount_or_warn(struct mount *mp)
5054 {
5055 	int error;
5056 
5057 	error = dounmount(mp, MNT_FORCE, curthread);
5058 	if (error != 0) {
5059 		printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
5060 		if (error == EBUSY)
5061 			printf("BUSY)\n");
5062 		else
5063 			printf("%d)\n", error);
5064 	}
5065 }
5066 
5067 /*
5068  * Unmount all filesystems. The list is traversed in reverse order
5069  * of mounting to avoid dependencies.
5070  */
5071 void
vfs_unmountall(void)5072 vfs_unmountall(void)
5073 {
5074 	struct mount *mp, *tmp;
5075 
5076 	CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
5077 
5078 	/*
5079 	 * Since this only runs when rebooting, it is not interlocked.
5080 	 */
5081 	TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
5082 		vfs_ref(mp);
5083 
5084 		/*
5085 		 * Forcibly unmounting "/dev" before "/" would prevent clean
5086 		 * unmount of the latter.
5087 		 */
5088 		if (mp == rootdevmp)
5089 			continue;
5090 
5091 		unmount_or_warn(mp);
5092 	}
5093 
5094 	if (rootdevmp != NULL)
5095 		unmount_or_warn(rootdevmp);
5096 }
5097 
5098 static void
vfs_deferred_inactive(struct vnode * vp,int lkflags)5099 vfs_deferred_inactive(struct vnode *vp, int lkflags)
5100 {
5101 
5102 	ASSERT_VI_LOCKED(vp, __func__);
5103 	VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
5104 	if ((vp->v_iflag & VI_OWEINACT) == 0) {
5105 		vdropl(vp);
5106 		return;
5107 	}
5108 	if (vn_lock(vp, lkflags) == 0) {
5109 		VI_LOCK(vp);
5110 		vinactive(vp);
5111 		VOP_UNLOCK(vp);
5112 		vdropl(vp);
5113 		return;
5114 	}
5115 	vdefer_inactive_unlocked(vp);
5116 }
5117 
5118 static int
vfs_periodic_inactive_filter(struct vnode * vp,void * arg)5119 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
5120 {
5121 
5122 	return (vp->v_iflag & VI_DEFINACT);
5123 }
5124 
5125 static void __noinline
vfs_periodic_inactive(struct mount * mp,int flags)5126 vfs_periodic_inactive(struct mount *mp, int flags)
5127 {
5128 	struct vnode *vp, *mvp;
5129 	int lkflags;
5130 
5131 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
5132 	if (flags != MNT_WAIT)
5133 		lkflags |= LK_NOWAIT;
5134 
5135 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
5136 		if ((vp->v_iflag & VI_DEFINACT) == 0) {
5137 			VI_UNLOCK(vp);
5138 			continue;
5139 		}
5140 		vp->v_iflag &= ~VI_DEFINACT;
5141 		vfs_deferred_inactive(vp, lkflags);
5142 	}
5143 }
5144 
5145 static inline bool
vfs_want_msync(struct vnode * vp)5146 vfs_want_msync(struct vnode *vp)
5147 {
5148 	struct vm_object *obj;
5149 
5150 	/*
5151 	 * This test may be performed without any locks held.
5152 	 * We rely on vm_object's type stability.
5153 	 */
5154 	if (vp->v_vflag & VV_NOSYNC)
5155 		return (false);
5156 	obj = vp->v_object;
5157 	return (obj != NULL && vm_object_mightbedirty(obj));
5158 }
5159 
5160 static int
vfs_periodic_msync_inactive_filter(struct vnode * vp,void * arg __unused)5161 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
5162 {
5163 
5164 	if (vp->v_vflag & VV_NOSYNC)
5165 		return (false);
5166 	if (vp->v_iflag & VI_DEFINACT)
5167 		return (true);
5168 	return (vfs_want_msync(vp));
5169 }
5170 
5171 static void __noinline
vfs_periodic_msync_inactive(struct mount * mp,int flags)5172 vfs_periodic_msync_inactive(struct mount *mp, int flags)
5173 {
5174 	struct vnode *vp, *mvp;
5175 	int lkflags;
5176 	bool seen_defer;
5177 
5178 	lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
5179 	if (flags != MNT_WAIT)
5180 		lkflags |= LK_NOWAIT;
5181 
5182 	MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
5183 		seen_defer = false;
5184 		if (vp->v_iflag & VI_DEFINACT) {
5185 			vp->v_iflag &= ~VI_DEFINACT;
5186 			seen_defer = true;
5187 		}
5188 		if (!vfs_want_msync(vp)) {
5189 			if (seen_defer)
5190 				vfs_deferred_inactive(vp, lkflags);
5191 			else
5192 				VI_UNLOCK(vp);
5193 			continue;
5194 		}
5195 		if (vget(vp, lkflags) == 0) {
5196 			if ((vp->v_vflag & VV_NOSYNC) == 0) {
5197 				if (flags == MNT_WAIT)
5198 					vnode_pager_clean_sync(vp);
5199 				else
5200 					vnode_pager_clean_async(vp);
5201 			}
5202 			vput(vp);
5203 			if (seen_defer)
5204 				vdrop(vp);
5205 		} else {
5206 			if (seen_defer)
5207 				vdefer_inactive_unlocked(vp);
5208 		}
5209 	}
5210 }
5211 
5212 void
vfs_periodic(struct mount * mp,int flags)5213 vfs_periodic(struct mount *mp, int flags)
5214 {
5215 
5216 	CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
5217 
5218 	if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
5219 		vfs_periodic_inactive(mp, flags);
5220 	else
5221 		vfs_periodic_msync_inactive(mp, flags);
5222 }
5223 
5224 static void
destroy_vpollinfo_free(struct vpollinfo * vi)5225 destroy_vpollinfo_free(struct vpollinfo *vi)
5226 {
5227 
5228 	knlist_destroy(&vi->vpi_selinfo.si_note);
5229 	mtx_destroy(&vi->vpi_lock);
5230 	free(vi, M_VNODEPOLL);
5231 }
5232 
5233 static void
destroy_vpollinfo(struct vpollinfo * vi)5234 destroy_vpollinfo(struct vpollinfo *vi)
5235 {
5236 
5237 	knlist_clear(&vi->vpi_selinfo.si_note, 1);
5238 	seldrain(&vi->vpi_selinfo);
5239 	destroy_vpollinfo_free(vi);
5240 }
5241 
5242 /*
5243  * Initialize per-vnode helper structure to hold poll-related state.
5244  */
5245 void
v_addpollinfo(struct vnode * vp)5246 v_addpollinfo(struct vnode *vp)
5247 {
5248 	struct vpollinfo *vi;
5249 
5250 	if (vp->v_pollinfo != NULL)
5251 		return;
5252 	vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
5253 	mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
5254 	knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
5255 	    vfs_knlunlock, vfs_knl_assert_lock);
5256 	VI_LOCK(vp);
5257 	if (vp->v_pollinfo != NULL) {
5258 		VI_UNLOCK(vp);
5259 		destroy_vpollinfo_free(vi);
5260 		return;
5261 	}
5262 	vp->v_pollinfo = vi;
5263 	VI_UNLOCK(vp);
5264 }
5265 
5266 /*
5267  * Record a process's interest in events which might happen to
5268  * a vnode.  Because poll uses the historic select-style interface
5269  * internally, this routine serves as both the ``check for any
5270  * pending events'' and the ``record my interest in future events''
5271  * functions.  (These are done together, while the lock is held,
5272  * to avoid race conditions.)
5273  */
5274 int
vn_pollrecord(struct vnode * vp,struct thread * td,int events)5275 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
5276 {
5277 
5278 	v_addpollinfo(vp);
5279 	mtx_lock(&vp->v_pollinfo->vpi_lock);
5280 	if (vp->v_pollinfo->vpi_revents & events) {
5281 		/*
5282 		 * This leaves events we are not interested
5283 		 * in available for the other process which
5284 		 * which presumably had requested them
5285 		 * (otherwise they would never have been
5286 		 * recorded).
5287 		 */
5288 		events &= vp->v_pollinfo->vpi_revents;
5289 		vp->v_pollinfo->vpi_revents &= ~events;
5290 
5291 		mtx_unlock(&vp->v_pollinfo->vpi_lock);
5292 		return (events);
5293 	}
5294 	vp->v_pollinfo->vpi_events |= events;
5295 	selrecord(td, &vp->v_pollinfo->vpi_selinfo);
5296 	mtx_unlock(&vp->v_pollinfo->vpi_lock);
5297 	return (0);
5298 }
5299 
5300 /*
5301  * Routine to create and manage a filesystem syncer vnode.
5302  */
5303 #define sync_close ((int (*)(struct  vop_close_args *))nullop)
5304 static int	sync_fsync(struct  vop_fsync_args *);
5305 static int	sync_inactive(struct  vop_inactive_args *);
5306 static int	sync_reclaim(struct  vop_reclaim_args *);
5307 
5308 static struct vop_vector sync_vnodeops = {
5309 	.vop_bypass =	VOP_EOPNOTSUPP,
5310 	.vop_close =	sync_close,
5311 	.vop_fsync =	sync_fsync,
5312 	.vop_getwritemount = vop_stdgetwritemount,
5313 	.vop_inactive =	sync_inactive,
5314 	.vop_need_inactive = vop_stdneed_inactive,
5315 	.vop_reclaim =	sync_reclaim,
5316 	.vop_lock1 =	vop_stdlock,
5317 	.vop_unlock =	vop_stdunlock,
5318 	.vop_islocked =	vop_stdislocked,
5319 	.vop_fplookup_vexec = VOP_EAGAIN,
5320 	.vop_fplookup_symlink = VOP_EAGAIN,
5321 };
5322 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
5323 
5324 /*
5325  * Create a new filesystem syncer vnode for the specified mount point.
5326  */
5327 void
vfs_allocate_syncvnode(struct mount * mp)5328 vfs_allocate_syncvnode(struct mount *mp)
5329 {
5330 	struct vnode *vp;
5331 	struct bufobj *bo;
5332 	static long start, incr, next;
5333 	int error;
5334 
5335 	/* Allocate a new vnode */
5336 	error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
5337 	if (error != 0)
5338 		panic("vfs_allocate_syncvnode: getnewvnode() failed");
5339 	vp->v_type = VNON;
5340 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5341 	vp->v_vflag |= VV_FORCEINSMQ;
5342 	error = insmntque1(vp, mp);
5343 	if (error != 0)
5344 		panic("vfs_allocate_syncvnode: insmntque() failed");
5345 	vp->v_vflag &= ~VV_FORCEINSMQ;
5346 	vn_set_state(vp, VSTATE_CONSTRUCTED);
5347 	VOP_UNLOCK(vp);
5348 	/*
5349 	 * Place the vnode onto the syncer worklist. We attempt to
5350 	 * scatter them about on the list so that they will go off
5351 	 * at evenly distributed times even if all the filesystems
5352 	 * are mounted at once.
5353 	 */
5354 	next += incr;
5355 	if (next == 0 || next > syncer_maxdelay) {
5356 		start /= 2;
5357 		incr /= 2;
5358 		if (start == 0) {
5359 			start = syncer_maxdelay / 2;
5360 			incr = syncer_maxdelay;
5361 		}
5362 		next = start;
5363 	}
5364 	bo = &vp->v_bufobj;
5365 	BO_LOCK(bo);
5366 	vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5367 	/* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5368 	mtx_lock(&sync_mtx);
5369 	sync_vnode_count++;
5370 	if (mp->mnt_syncer == NULL) {
5371 		mp->mnt_syncer = vp;
5372 		vp = NULL;
5373 	}
5374 	mtx_unlock(&sync_mtx);
5375 	BO_UNLOCK(bo);
5376 	if (vp != NULL) {
5377 		vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5378 		vgone(vp);
5379 		vput(vp);
5380 	}
5381 }
5382 
5383 void
vfs_deallocate_syncvnode(struct mount * mp)5384 vfs_deallocate_syncvnode(struct mount *mp)
5385 {
5386 	struct vnode *vp;
5387 
5388 	mtx_lock(&sync_mtx);
5389 	vp = mp->mnt_syncer;
5390 	if (vp != NULL)
5391 		mp->mnt_syncer = NULL;
5392 	mtx_unlock(&sync_mtx);
5393 	if (vp != NULL)
5394 		vrele(vp);
5395 }
5396 
5397 /*
5398  * Do a lazy sync of the filesystem.
5399  */
5400 static int
sync_fsync(struct vop_fsync_args * ap)5401 sync_fsync(struct vop_fsync_args *ap)
5402 {
5403 	struct vnode *syncvp = ap->a_vp;
5404 	struct mount *mp = syncvp->v_mount;
5405 	int error, save;
5406 	struct bufobj *bo;
5407 
5408 	/*
5409 	 * We only need to do something if this is a lazy evaluation.
5410 	 */
5411 	if (ap->a_waitfor != MNT_LAZY)
5412 		return (0);
5413 
5414 	/*
5415 	 * Move ourselves to the back of the sync list.
5416 	 */
5417 	bo = &syncvp->v_bufobj;
5418 	BO_LOCK(bo);
5419 	vn_syncer_add_to_worklist(bo, syncdelay);
5420 	BO_UNLOCK(bo);
5421 
5422 	/*
5423 	 * Walk the list of vnodes pushing all that are dirty and
5424 	 * not already on the sync list.
5425 	 */
5426 	if (vfs_busy(mp, MBF_NOWAIT) != 0)
5427 		return (0);
5428 	VOP_UNLOCK(syncvp);
5429 	save = curthread_pflags_set(TDP_SYNCIO);
5430 	/*
5431 	 * The filesystem at hand may be idle with free vnodes stored in the
5432 	 * batch.  Return them instead of letting them stay there indefinitely.
5433 	 */
5434 	vfs_periodic(mp, MNT_NOWAIT);
5435 	error = VFS_SYNC(mp, MNT_LAZY);
5436 	curthread_pflags_restore(save);
5437 	vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY);
5438 	vfs_unbusy(mp);
5439 	return (error);
5440 }
5441 
5442 /*
5443  * The syncer vnode is no referenced.
5444  */
5445 static int
sync_inactive(struct vop_inactive_args * ap)5446 sync_inactive(struct vop_inactive_args *ap)
5447 {
5448 
5449 	vgone(ap->a_vp);
5450 	return (0);
5451 }
5452 
5453 /*
5454  * The syncer vnode is no longer needed and is being decommissioned.
5455  *
5456  * Modifications to the worklist must be protected by sync_mtx.
5457  */
5458 static int
sync_reclaim(struct vop_reclaim_args * ap)5459 sync_reclaim(struct vop_reclaim_args *ap)
5460 {
5461 	struct vnode *vp = ap->a_vp;
5462 	struct bufobj *bo;
5463 
5464 	bo = &vp->v_bufobj;
5465 	BO_LOCK(bo);
5466 	mtx_lock(&sync_mtx);
5467 	if (vp->v_mount->mnt_syncer == vp)
5468 		vp->v_mount->mnt_syncer = NULL;
5469 	if (bo->bo_flag & BO_ONWORKLST) {
5470 		LIST_REMOVE(bo, bo_synclist);
5471 		syncer_worklist_len--;
5472 		sync_vnode_count--;
5473 		bo->bo_flag &= ~BO_ONWORKLST;
5474 	}
5475 	mtx_unlock(&sync_mtx);
5476 	BO_UNLOCK(bo);
5477 
5478 	return (0);
5479 }
5480 
5481 int
vn_need_pageq_flush(struct vnode * vp)5482 vn_need_pageq_flush(struct vnode *vp)
5483 {
5484 	struct vm_object *obj;
5485 
5486 	obj = vp->v_object;
5487 	return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5488 	    vm_object_mightbedirty(obj));
5489 }
5490 
5491 /*
5492  * Check if vnode represents a disk device
5493  */
5494 bool
vn_isdisk_error(struct vnode * vp,int * errp)5495 vn_isdisk_error(struct vnode *vp, int *errp)
5496 {
5497 	int error;
5498 
5499 	if (vp->v_type != VCHR) {
5500 		error = ENOTBLK;
5501 		goto out;
5502 	}
5503 	error = 0;
5504 	dev_lock();
5505 	if (vp->v_rdev == NULL)
5506 		error = ENXIO;
5507 	else if (vp->v_rdev->si_devsw == NULL)
5508 		error = ENXIO;
5509 	else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5510 		error = ENOTBLK;
5511 	dev_unlock();
5512 out:
5513 	*errp = error;
5514 	return (error == 0);
5515 }
5516 
5517 bool
vn_isdisk(struct vnode * vp)5518 vn_isdisk(struct vnode *vp)
5519 {
5520 	int error;
5521 
5522 	return (vn_isdisk_error(vp, &error));
5523 }
5524 
5525 /*
5526  * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5527  * the comment above cache_fplookup for details.
5528  */
5529 int
vaccess_vexec_smr(mode_t file_mode,uid_t file_uid,gid_t file_gid,struct ucred * cred)5530 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5531 {
5532 	int error;
5533 
5534 	VFS_SMR_ASSERT_ENTERED();
5535 
5536 	/* Check the owner. */
5537 	if (cred->cr_uid == file_uid) {
5538 		if (file_mode & S_IXUSR)
5539 			return (0);
5540 		goto out_error;
5541 	}
5542 
5543 	/* Otherwise, check the groups (first match) */
5544 	if (groupmember(file_gid, cred)) {
5545 		if (file_mode & S_IXGRP)
5546 			return (0);
5547 		goto out_error;
5548 	}
5549 
5550 	/* Otherwise, check everyone else. */
5551 	if (file_mode & S_IXOTH)
5552 		return (0);
5553 out_error:
5554 	/*
5555 	 * Permission check failed, but it is possible denial will get overwritten
5556 	 * (e.g., when root is traversing through a 700 directory owned by someone
5557 	 * else).
5558 	 *
5559 	 * vaccess() calls priv_check_cred which in turn can descent into MAC
5560 	 * modules overriding this result. It's quite unclear what semantics
5561 	 * are allowed for them to operate, thus for safety we don't call them
5562 	 * from within the SMR section. This also means if any such modules
5563 	 * are present, we have to let the regular lookup decide.
5564 	 */
5565 	error = priv_check_cred_vfs_lookup_nomac(cred);
5566 	switch (error) {
5567 	case 0:
5568 		return (0);
5569 	case EAGAIN:
5570 		/*
5571 		 * MAC modules present.
5572 		 */
5573 		return (EAGAIN);
5574 	case EPERM:
5575 		return (EACCES);
5576 	default:
5577 		return (error);
5578 	}
5579 }
5580 
5581 /*
5582  * Common filesystem object access control check routine.  Accepts a
5583  * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5584  * Returns 0 on success, or an errno on failure.
5585  */
5586 int
vaccess(__enum_uint8 (vtype)type,mode_t file_mode,uid_t file_uid,gid_t file_gid,accmode_t accmode,struct ucred * cred)5587 vaccess(__enum_uint8(vtype) type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5588     accmode_t accmode, struct ucred *cred)
5589 {
5590 	accmode_t dac_granted;
5591 	accmode_t priv_granted;
5592 
5593 	KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5594 	    ("invalid bit in accmode"));
5595 	KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5596 	    ("VAPPEND without VWRITE"));
5597 
5598 	/*
5599 	 * Look for a normal, non-privileged way to access the file/directory
5600 	 * as requested.  If it exists, go with that.
5601 	 */
5602 
5603 	dac_granted = 0;
5604 
5605 	/* Check the owner. */
5606 	if (cred->cr_uid == file_uid) {
5607 		dac_granted |= VADMIN;
5608 		if (file_mode & S_IXUSR)
5609 			dac_granted |= VEXEC;
5610 		if (file_mode & S_IRUSR)
5611 			dac_granted |= VREAD;
5612 		if (file_mode & S_IWUSR)
5613 			dac_granted |= (VWRITE | VAPPEND);
5614 
5615 		if ((accmode & dac_granted) == accmode)
5616 			return (0);
5617 
5618 		goto privcheck;
5619 	}
5620 
5621 	/* Otherwise, check the groups (first match) */
5622 	if (groupmember(file_gid, cred)) {
5623 		if (file_mode & S_IXGRP)
5624 			dac_granted |= VEXEC;
5625 		if (file_mode & S_IRGRP)
5626 			dac_granted |= VREAD;
5627 		if (file_mode & S_IWGRP)
5628 			dac_granted |= (VWRITE | VAPPEND);
5629 
5630 		if ((accmode & dac_granted) == accmode)
5631 			return (0);
5632 
5633 		goto privcheck;
5634 	}
5635 
5636 	/* Otherwise, check everyone else. */
5637 	if (file_mode & S_IXOTH)
5638 		dac_granted |= VEXEC;
5639 	if (file_mode & S_IROTH)
5640 		dac_granted |= VREAD;
5641 	if (file_mode & S_IWOTH)
5642 		dac_granted |= (VWRITE | VAPPEND);
5643 	if ((accmode & dac_granted) == accmode)
5644 		return (0);
5645 
5646 privcheck:
5647 	/*
5648 	 * Build a privilege mask to determine if the set of privileges
5649 	 * satisfies the requirements when combined with the granted mask
5650 	 * from above.  For each privilege, if the privilege is required,
5651 	 * bitwise or the request type onto the priv_granted mask.
5652 	 */
5653 	priv_granted = 0;
5654 
5655 	if (type == VDIR) {
5656 		/*
5657 		 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5658 		 * requests, instead of PRIV_VFS_EXEC.
5659 		 */
5660 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5661 		    !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5662 			priv_granted |= VEXEC;
5663 	} else {
5664 		/*
5665 		 * Ensure that at least one execute bit is on. Otherwise,
5666 		 * a privileged user will always succeed, and we don't want
5667 		 * this to happen unless the file really is executable.
5668 		 */
5669 		if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5670 		    (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5671 		    !priv_check_cred(cred, PRIV_VFS_EXEC))
5672 			priv_granted |= VEXEC;
5673 	}
5674 
5675 	if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5676 	    !priv_check_cred(cred, PRIV_VFS_READ))
5677 		priv_granted |= VREAD;
5678 
5679 	if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5680 	    !priv_check_cred(cred, PRIV_VFS_WRITE))
5681 		priv_granted |= (VWRITE | VAPPEND);
5682 
5683 	if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5684 	    !priv_check_cred(cred, PRIV_VFS_ADMIN))
5685 		priv_granted |= VADMIN;
5686 
5687 	if ((accmode & (priv_granted | dac_granted)) == accmode) {
5688 		return (0);
5689 	}
5690 
5691 	return ((accmode & VADMIN) ? EPERM : EACCES);
5692 }
5693 
5694 /*
5695  * Credential check based on process requesting service, and per-attribute
5696  * permissions.
5697  */
5698 int
extattr_check_cred(struct vnode * vp,int attrnamespace,struct ucred * cred,struct thread * td,accmode_t accmode)5699 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5700     struct thread *td, accmode_t accmode)
5701 {
5702 
5703 	/*
5704 	 * Kernel-invoked always succeeds.
5705 	 */
5706 	if (cred == NOCRED)
5707 		return (0);
5708 
5709 	/*
5710 	 * Do not allow privileged processes in jail to directly manipulate
5711 	 * system attributes.
5712 	 */
5713 	switch (attrnamespace) {
5714 	case EXTATTR_NAMESPACE_SYSTEM:
5715 		/* Potentially should be: return (EPERM); */
5716 		return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5717 	case EXTATTR_NAMESPACE_USER:
5718 		return (VOP_ACCESS(vp, accmode, cred, td));
5719 	default:
5720 		return (EPERM);
5721 	}
5722 }
5723 
5724 #ifdef DEBUG_VFS_LOCKS
5725 int vfs_badlock_ddb = 1;	/* Drop into debugger on violation. */
5726 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5727     "Drop into debugger on lock violation");
5728 
5729 int vfs_badlock_mutex = 1;	/* Check for interlock across VOPs. */
5730 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5731     0, "Check for interlock across VOPs");
5732 
5733 int vfs_badlock_print = 1;	/* Print lock violations. */
5734 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5735     0, "Print lock violations");
5736 
5737 int vfs_badlock_vnode = 1;	/* Print vnode details on lock violations. */
5738 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5739     0, "Print vnode details on lock violations");
5740 
5741 #ifdef KDB
5742 int vfs_badlock_backtrace = 1;	/* Print backtrace at lock violations. */
5743 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5744     &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5745 #endif
5746 
5747 static void
vfs_badlock(const char * msg,const char * str,struct vnode * vp)5748 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5749 {
5750 
5751 #ifdef KDB
5752 	if (vfs_badlock_backtrace)
5753 		kdb_backtrace();
5754 #endif
5755 	if (vfs_badlock_vnode)
5756 		vn_printf(vp, "vnode ");
5757 	if (vfs_badlock_print)
5758 		printf("%s: %p %s\n", str, (void *)vp, msg);
5759 	if (vfs_badlock_ddb)
5760 		kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5761 }
5762 
5763 void
assert_vi_locked(struct vnode * vp,const char * str)5764 assert_vi_locked(struct vnode *vp, const char *str)
5765 {
5766 
5767 	if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5768 		vfs_badlock("interlock is not locked but should be", str, vp);
5769 }
5770 
5771 void
assert_vi_unlocked(struct vnode * vp,const char * str)5772 assert_vi_unlocked(struct vnode *vp, const char *str)
5773 {
5774 
5775 	if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5776 		vfs_badlock("interlock is locked but should not be", str, vp);
5777 }
5778 
5779 void
assert_vop_locked(struct vnode * vp,const char * str)5780 assert_vop_locked(struct vnode *vp, const char *str)
5781 {
5782 	if (KERNEL_PANICKED() || vp == NULL)
5783 		return;
5784 
5785 #ifdef WITNESS
5786 	if ((vp->v_irflag & VIRF_CROSSMP) == 0 &&
5787 	    witness_is_owned(&vp->v_vnlock->lock_object) == -1)
5788 #else
5789 	int locked = VOP_ISLOCKED(vp);
5790 	if (locked == 0 || locked == LK_EXCLOTHER)
5791 #endif
5792 		vfs_badlock("is not locked but should be", str, vp);
5793 }
5794 
5795 void
assert_vop_unlocked(struct vnode * vp,const char * str)5796 assert_vop_unlocked(struct vnode *vp, const char *str)
5797 {
5798 	if (KERNEL_PANICKED() || vp == NULL)
5799 		return;
5800 
5801 #ifdef WITNESS
5802 	if ((vp->v_irflag & VIRF_CROSSMP) == 0 &&
5803 	    witness_is_owned(&vp->v_vnlock->lock_object) == 1)
5804 #else
5805 	if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5806 #endif
5807 		vfs_badlock("is locked but should not be", str, vp);
5808 }
5809 
5810 void
assert_vop_elocked(struct vnode * vp,const char * str)5811 assert_vop_elocked(struct vnode *vp, const char *str)
5812 {
5813 	if (KERNEL_PANICKED() || vp == NULL)
5814 		return;
5815 
5816 	if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5817 		vfs_badlock("is not exclusive locked but should be", str, vp);
5818 }
5819 #endif /* DEBUG_VFS_LOCKS */
5820 
5821 void
vop_rename_fail(struct vop_rename_args * ap)5822 vop_rename_fail(struct vop_rename_args *ap)
5823 {
5824 
5825 	if (ap->a_tvp != NULL)
5826 		vput(ap->a_tvp);
5827 	if (ap->a_tdvp == ap->a_tvp)
5828 		vrele(ap->a_tdvp);
5829 	else
5830 		vput(ap->a_tdvp);
5831 	vrele(ap->a_fdvp);
5832 	vrele(ap->a_fvp);
5833 }
5834 
5835 void
vop_rename_pre(void * ap)5836 vop_rename_pre(void *ap)
5837 {
5838 	struct vop_rename_args *a = ap;
5839 
5840 #ifdef DEBUG_VFS_LOCKS
5841 	if (a->a_tvp)
5842 		ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5843 	ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5844 	ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5845 	ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5846 
5847 	/* Check the source (from). */
5848 	if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5849 	    (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5850 		ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5851 	if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5852 		ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5853 
5854 	/* Check the target. */
5855 	if (a->a_tvp)
5856 		ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5857 	ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5858 #endif
5859 	/*
5860 	 * It may be tempting to add vn_seqc_write_begin/end calls here and
5861 	 * in vop_rename_post but that's not going to work out since some
5862 	 * filesystems relookup vnodes mid-rename. This is probably a bug.
5863 	 *
5864 	 * For now filesystems are expected to do the relevant calls after they
5865 	 * decide what vnodes to operate on.
5866 	 */
5867 	if (a->a_tdvp != a->a_fdvp)
5868 		vhold(a->a_fdvp);
5869 	if (a->a_tvp != a->a_fvp)
5870 		vhold(a->a_fvp);
5871 	vhold(a->a_tdvp);
5872 	if (a->a_tvp)
5873 		vhold(a->a_tvp);
5874 }
5875 
5876 #ifdef DEBUG_VFS_LOCKS
5877 void
vop_fplookup_vexec_debugpre(void * ap __unused)5878 vop_fplookup_vexec_debugpre(void *ap __unused)
5879 {
5880 
5881 	VFS_SMR_ASSERT_ENTERED();
5882 }
5883 
5884 void
vop_fplookup_vexec_debugpost(void * ap,int rc)5885 vop_fplookup_vexec_debugpost(void *ap, int rc)
5886 {
5887 	struct vop_fplookup_vexec_args *a;
5888 	struct vnode *vp;
5889 
5890 	a = ap;
5891 	vp = a->a_vp;
5892 
5893 	VFS_SMR_ASSERT_ENTERED();
5894 	if (rc == EOPNOTSUPP)
5895 		VNPASS(VN_IS_DOOMED(vp), vp);
5896 }
5897 
5898 void
vop_fplookup_symlink_debugpre(void * ap __unused)5899 vop_fplookup_symlink_debugpre(void *ap __unused)
5900 {
5901 
5902 	VFS_SMR_ASSERT_ENTERED();
5903 }
5904 
5905 void
vop_fplookup_symlink_debugpost(void * ap __unused,int rc __unused)5906 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5907 {
5908 
5909 	VFS_SMR_ASSERT_ENTERED();
5910 }
5911 
5912 static void
vop_fsync_debugprepost(struct vnode * vp,const char * name)5913 vop_fsync_debugprepost(struct vnode *vp, const char *name)
5914 {
5915 	if (vp->v_type == VCHR)
5916 		;
5917 	/*
5918 	 * The shared vs. exclusive locking policy for fsync()
5919 	 * is actually determined by vp's write mount as indicated
5920 	 * by VOP_GETWRITEMOUNT(), which for stacked filesystems
5921 	 * may not be the same as vp->v_mount.  However, if the
5922 	 * underlying filesystem which really handles the fsync()
5923 	 * supports shared locking, the stacked filesystem must also
5924 	 * be prepared for its VOP_FSYNC() operation to be called
5925 	 * with only a shared lock.  On the other hand, if the
5926 	 * stacked filesystem claims support for shared write
5927 	 * locking but the underlying filesystem does not, and the
5928 	 * caller incorrectly uses a shared lock, this condition
5929 	 * should still be caught when the stacked filesystem
5930 	 * invokes VOP_FSYNC() on the underlying filesystem.
5931 	 */
5932 	else if (MNT_SHARED_WRITES(vp->v_mount))
5933 		ASSERT_VOP_LOCKED(vp, name);
5934 	else
5935 		ASSERT_VOP_ELOCKED(vp, name);
5936 }
5937 
5938 void
vop_fsync_debugpre(void * a)5939 vop_fsync_debugpre(void *a)
5940 {
5941 	struct vop_fsync_args *ap;
5942 
5943 	ap = a;
5944 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5945 }
5946 
5947 void
vop_fsync_debugpost(void * a,int rc __unused)5948 vop_fsync_debugpost(void *a, int rc __unused)
5949 {
5950 	struct vop_fsync_args *ap;
5951 
5952 	ap = a;
5953 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5954 }
5955 
5956 void
vop_fdatasync_debugpre(void * a)5957 vop_fdatasync_debugpre(void *a)
5958 {
5959 	struct vop_fdatasync_args *ap;
5960 
5961 	ap = a;
5962 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5963 }
5964 
5965 void
vop_fdatasync_debugpost(void * a,int rc __unused)5966 vop_fdatasync_debugpost(void *a, int rc __unused)
5967 {
5968 	struct vop_fdatasync_args *ap;
5969 
5970 	ap = a;
5971 	vop_fsync_debugprepost(ap->a_vp, "fsync");
5972 }
5973 
5974 void
vop_strategy_debugpre(void * ap)5975 vop_strategy_debugpre(void *ap)
5976 {
5977 	struct vop_strategy_args *a;
5978 	struct buf *bp;
5979 
5980 	a = ap;
5981 	bp = a->a_bp;
5982 
5983 	/*
5984 	 * Cluster ops lock their component buffers but not the IO container.
5985 	 */
5986 	if ((bp->b_flags & B_CLUSTER) != 0)
5987 		return;
5988 
5989 	if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5990 		if (vfs_badlock_print)
5991 			printf(
5992 			    "VOP_STRATEGY: bp is not locked but should be\n");
5993 		if (vfs_badlock_ddb)
5994 			kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5995 	}
5996 }
5997 
5998 void
vop_lock_debugpre(void * ap)5999 vop_lock_debugpre(void *ap)
6000 {
6001 	struct vop_lock1_args *a = ap;
6002 
6003 	if ((a->a_flags & LK_INTERLOCK) == 0)
6004 		ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
6005 	else
6006 		ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
6007 }
6008 
6009 void
vop_lock_debugpost(void * ap,int rc)6010 vop_lock_debugpost(void *ap, int rc)
6011 {
6012 	struct vop_lock1_args *a = ap;
6013 
6014 	ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
6015 	if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
6016 		ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
6017 }
6018 
6019 void
vop_unlock_debugpre(void * ap)6020 vop_unlock_debugpre(void *ap)
6021 {
6022 	struct vop_unlock_args *a = ap;
6023 	struct vnode *vp = a->a_vp;
6024 
6025 	VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp);
6026 	ASSERT_VOP_LOCKED(vp, "VOP_UNLOCK");
6027 }
6028 
6029 void
vop_need_inactive_debugpre(void * ap)6030 vop_need_inactive_debugpre(void *ap)
6031 {
6032 	struct vop_need_inactive_args *a = ap;
6033 
6034 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
6035 }
6036 
6037 void
vop_need_inactive_debugpost(void * ap,int rc)6038 vop_need_inactive_debugpost(void *ap, int rc)
6039 {
6040 	struct vop_need_inactive_args *a = ap;
6041 
6042 	ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
6043 }
6044 #endif
6045 
6046 void
vop_create_pre(void * ap)6047 vop_create_pre(void *ap)
6048 {
6049 	struct vop_create_args *a;
6050 	struct vnode *dvp;
6051 
6052 	a = ap;
6053 	dvp = a->a_dvp;
6054 	vn_seqc_write_begin(dvp);
6055 }
6056 
6057 void
vop_create_post(void * ap,int rc)6058 vop_create_post(void *ap, int rc)
6059 {
6060 	struct vop_create_args *a;
6061 	struct vnode *dvp;
6062 
6063 	a = ap;
6064 	dvp = a->a_dvp;
6065 	vn_seqc_write_end(dvp);
6066 	if (!rc)
6067 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6068 }
6069 
6070 void
vop_whiteout_pre(void * ap)6071 vop_whiteout_pre(void *ap)
6072 {
6073 	struct vop_whiteout_args *a;
6074 	struct vnode *dvp;
6075 
6076 	a = ap;
6077 	dvp = a->a_dvp;
6078 	vn_seqc_write_begin(dvp);
6079 }
6080 
6081 void
vop_whiteout_post(void * ap,int rc)6082 vop_whiteout_post(void *ap, int rc)
6083 {
6084 	struct vop_whiteout_args *a;
6085 	struct vnode *dvp;
6086 
6087 	a = ap;
6088 	dvp = a->a_dvp;
6089 	vn_seqc_write_end(dvp);
6090 }
6091 
6092 void
vop_deleteextattr_pre(void * ap)6093 vop_deleteextattr_pre(void *ap)
6094 {
6095 	struct vop_deleteextattr_args *a;
6096 	struct vnode *vp;
6097 
6098 	a = ap;
6099 	vp = a->a_vp;
6100 	vn_seqc_write_begin(vp);
6101 }
6102 
6103 void
vop_deleteextattr_post(void * ap,int rc)6104 vop_deleteextattr_post(void *ap, int rc)
6105 {
6106 	struct vop_deleteextattr_args *a;
6107 	struct vnode *vp;
6108 
6109 	a = ap;
6110 	vp = a->a_vp;
6111 	vn_seqc_write_end(vp);
6112 	if (!rc)
6113 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
6114 }
6115 
6116 void
vop_link_pre(void * ap)6117 vop_link_pre(void *ap)
6118 {
6119 	struct vop_link_args *a;
6120 	struct vnode *vp, *tdvp;
6121 
6122 	a = ap;
6123 	vp = a->a_vp;
6124 	tdvp = a->a_tdvp;
6125 	vn_seqc_write_begin(vp);
6126 	vn_seqc_write_begin(tdvp);
6127 }
6128 
6129 void
vop_link_post(void * ap,int rc)6130 vop_link_post(void *ap, int rc)
6131 {
6132 	struct vop_link_args *a;
6133 	struct vnode *vp, *tdvp;
6134 
6135 	a = ap;
6136 	vp = a->a_vp;
6137 	tdvp = a->a_tdvp;
6138 	vn_seqc_write_end(vp);
6139 	vn_seqc_write_end(tdvp);
6140 	if (!rc) {
6141 		VFS_KNOTE_LOCKED(vp, NOTE_LINK);
6142 		VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
6143 	}
6144 }
6145 
6146 void
vop_mkdir_pre(void * ap)6147 vop_mkdir_pre(void *ap)
6148 {
6149 	struct vop_mkdir_args *a;
6150 	struct vnode *dvp;
6151 
6152 	a = ap;
6153 	dvp = a->a_dvp;
6154 	vn_seqc_write_begin(dvp);
6155 }
6156 
6157 void
vop_mkdir_post(void * ap,int rc)6158 vop_mkdir_post(void *ap, int rc)
6159 {
6160 	struct vop_mkdir_args *a;
6161 	struct vnode *dvp;
6162 
6163 	a = ap;
6164 	dvp = a->a_dvp;
6165 	vn_seqc_write_end(dvp);
6166 	if (!rc)
6167 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
6168 }
6169 
6170 #ifdef DEBUG_VFS_LOCKS
6171 void
vop_mkdir_debugpost(void * ap,int rc)6172 vop_mkdir_debugpost(void *ap, int rc)
6173 {
6174 	struct vop_mkdir_args *a;
6175 
6176 	a = ap;
6177 	if (!rc)
6178 		cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
6179 }
6180 #endif
6181 
6182 void
vop_mknod_pre(void * ap)6183 vop_mknod_pre(void *ap)
6184 {
6185 	struct vop_mknod_args *a;
6186 	struct vnode *dvp;
6187 
6188 	a = ap;
6189 	dvp = a->a_dvp;
6190 	vn_seqc_write_begin(dvp);
6191 }
6192 
6193 void
vop_mknod_post(void * ap,int rc)6194 vop_mknod_post(void *ap, int rc)
6195 {
6196 	struct vop_mknod_args *a;
6197 	struct vnode *dvp;
6198 
6199 	a = ap;
6200 	dvp = a->a_dvp;
6201 	vn_seqc_write_end(dvp);
6202 	if (!rc)
6203 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6204 }
6205 
6206 void
vop_reclaim_post(void * ap,int rc)6207 vop_reclaim_post(void *ap, int rc)
6208 {
6209 	struct vop_reclaim_args *a;
6210 	struct vnode *vp;
6211 
6212 	a = ap;
6213 	vp = a->a_vp;
6214 	ASSERT_VOP_IN_SEQC(vp);
6215 	if (!rc)
6216 		VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
6217 }
6218 
6219 void
vop_remove_pre(void * ap)6220 vop_remove_pre(void *ap)
6221 {
6222 	struct vop_remove_args *a;
6223 	struct vnode *dvp, *vp;
6224 
6225 	a = ap;
6226 	dvp = a->a_dvp;
6227 	vp = a->a_vp;
6228 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_UNLINK);
6229 	vn_seqc_write_begin(dvp);
6230 	vn_seqc_write_begin(vp);
6231 }
6232 
6233 void
vop_remove_post(void * ap,int rc)6234 vop_remove_post(void *ap, int rc)
6235 {
6236 	struct vop_remove_args *a;
6237 	struct vnode *dvp, *vp;
6238 
6239 	a = ap;
6240 	dvp = a->a_dvp;
6241 	vp = a->a_vp;
6242 	vn_seqc_write_end(dvp);
6243 	vn_seqc_write_end(vp);
6244 	if (!rc) {
6245 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6246 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
6247 	}
6248 }
6249 
6250 void
vop_rename_post(void * ap,int rc)6251 vop_rename_post(void *ap, int rc)
6252 {
6253 	struct vop_rename_args *a = ap;
6254 	long hint;
6255 
6256 	if (!rc) {
6257 		hint = NOTE_WRITE;
6258 		if (a->a_fdvp == a->a_tdvp) {
6259 			if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
6260 				hint |= NOTE_LINK;
6261 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
6262 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
6263 		} else {
6264 			hint |= NOTE_EXTEND;
6265 			if (a->a_fvp->v_type == VDIR)
6266 				hint |= NOTE_LINK;
6267 			VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
6268 
6269 			if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
6270 			    a->a_tvp->v_type == VDIR)
6271 				hint &= ~NOTE_LINK;
6272 			VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
6273 		}
6274 
6275 		VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
6276 		if (a->a_tvp)
6277 			VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
6278 	}
6279 	if (a->a_tdvp != a->a_fdvp)
6280 		vdrop(a->a_fdvp);
6281 	if (a->a_tvp != a->a_fvp)
6282 		vdrop(a->a_fvp);
6283 	vdrop(a->a_tdvp);
6284 	if (a->a_tvp)
6285 		vdrop(a->a_tvp);
6286 }
6287 
6288 void
vop_rmdir_pre(void * ap)6289 vop_rmdir_pre(void *ap)
6290 {
6291 	struct vop_rmdir_args *a;
6292 	struct vnode *dvp, *vp;
6293 
6294 	a = ap;
6295 	dvp = a->a_dvp;
6296 	vp = a->a_vp;
6297 	vfs_notify_upper(vp, VFS_NOTIFY_UPPER_UNLINK);
6298 	vn_seqc_write_begin(dvp);
6299 	vn_seqc_write_begin(vp);
6300 }
6301 
6302 void
vop_rmdir_post(void * ap,int rc)6303 vop_rmdir_post(void *ap, int rc)
6304 {
6305 	struct vop_rmdir_args *a;
6306 	struct vnode *dvp, *vp;
6307 
6308 	a = ap;
6309 	dvp = a->a_dvp;
6310 	vp = a->a_vp;
6311 	vn_seqc_write_end(dvp);
6312 	vn_seqc_write_end(vp);
6313 	if (!rc) {
6314 		vp->v_vflag |= VV_UNLINKED;
6315 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
6316 		VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
6317 	}
6318 }
6319 
6320 void
vop_setattr_pre(void * ap)6321 vop_setattr_pre(void *ap)
6322 {
6323 	struct vop_setattr_args *a;
6324 	struct vnode *vp;
6325 
6326 	a = ap;
6327 	vp = a->a_vp;
6328 	vn_seqc_write_begin(vp);
6329 }
6330 
6331 void
vop_setattr_post(void * ap,int rc)6332 vop_setattr_post(void *ap, int rc)
6333 {
6334 	struct vop_setattr_args *a;
6335 	struct vnode *vp;
6336 
6337 	a = ap;
6338 	vp = a->a_vp;
6339 	vn_seqc_write_end(vp);
6340 	if (!rc)
6341 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6342 }
6343 
6344 void
vop_setacl_pre(void * ap)6345 vop_setacl_pre(void *ap)
6346 {
6347 	struct vop_setacl_args *a;
6348 	struct vnode *vp;
6349 
6350 	a = ap;
6351 	vp = a->a_vp;
6352 	vn_seqc_write_begin(vp);
6353 }
6354 
6355 void
vop_setacl_post(void * ap,int rc __unused)6356 vop_setacl_post(void *ap, int rc __unused)
6357 {
6358 	struct vop_setacl_args *a;
6359 	struct vnode *vp;
6360 
6361 	a = ap;
6362 	vp = a->a_vp;
6363 	vn_seqc_write_end(vp);
6364 }
6365 
6366 void
vop_setextattr_pre(void * ap)6367 vop_setextattr_pre(void *ap)
6368 {
6369 	struct vop_setextattr_args *a;
6370 	struct vnode *vp;
6371 
6372 	a = ap;
6373 	vp = a->a_vp;
6374 	vn_seqc_write_begin(vp);
6375 }
6376 
6377 void
vop_setextattr_post(void * ap,int rc)6378 vop_setextattr_post(void *ap, int rc)
6379 {
6380 	struct vop_setextattr_args *a;
6381 	struct vnode *vp;
6382 
6383 	a = ap;
6384 	vp = a->a_vp;
6385 	vn_seqc_write_end(vp);
6386 	if (!rc)
6387 		VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6388 }
6389 
6390 void
vop_symlink_pre(void * ap)6391 vop_symlink_pre(void *ap)
6392 {
6393 	struct vop_symlink_args *a;
6394 	struct vnode *dvp;
6395 
6396 	a = ap;
6397 	dvp = a->a_dvp;
6398 	vn_seqc_write_begin(dvp);
6399 }
6400 
6401 void
vop_symlink_post(void * ap,int rc)6402 vop_symlink_post(void *ap, int rc)
6403 {
6404 	struct vop_symlink_args *a;
6405 	struct vnode *dvp;
6406 
6407 	a = ap;
6408 	dvp = a->a_dvp;
6409 	vn_seqc_write_end(dvp);
6410 	if (!rc)
6411 		VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6412 }
6413 
6414 void
vop_open_post(void * ap,int rc)6415 vop_open_post(void *ap, int rc)
6416 {
6417 	struct vop_open_args *a = ap;
6418 
6419 	if (!rc)
6420 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6421 }
6422 
6423 void
vop_close_post(void * ap,int rc)6424 vop_close_post(void *ap, int rc)
6425 {
6426 	struct vop_close_args *a = ap;
6427 
6428 	if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6429 	    !VN_IS_DOOMED(a->a_vp))) {
6430 		VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6431 		    NOTE_CLOSE_WRITE : NOTE_CLOSE);
6432 	}
6433 }
6434 
6435 void
vop_read_post(void * ap,int rc)6436 vop_read_post(void *ap, int rc)
6437 {
6438 	struct vop_read_args *a = ap;
6439 
6440 	if (!rc)
6441 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6442 }
6443 
6444 void
vop_read_pgcache_post(void * ap,int rc)6445 vop_read_pgcache_post(void *ap, int rc)
6446 {
6447 	struct vop_read_pgcache_args *a = ap;
6448 
6449 	if (!rc)
6450 		VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6451 }
6452 
6453 void
vop_readdir_post(void * ap,int rc)6454 vop_readdir_post(void *ap, int rc)
6455 {
6456 	struct vop_readdir_args *a = ap;
6457 
6458 	if (!rc)
6459 		VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6460 }
6461 
6462 static struct knlist fs_knlist;
6463 
6464 static void
vfs_event_init(void * arg)6465 vfs_event_init(void *arg)
6466 {
6467 	knlist_init_mtx(&fs_knlist, NULL);
6468 }
6469 /* XXX - correct order? */
6470 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6471 
6472 void
vfs_event_signal(fsid_t * fsid,uint32_t event,intptr_t data __unused)6473 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6474 {
6475 
6476 	KNOTE_UNLOCKED(&fs_knlist, event);
6477 }
6478 
6479 static int	filt_fsattach(struct knote *kn);
6480 static void	filt_fsdetach(struct knote *kn);
6481 static int	filt_fsevent(struct knote *kn, long hint);
6482 
6483 const struct filterops fs_filtops = {
6484 	.f_isfd = 0,
6485 	.f_attach = filt_fsattach,
6486 	.f_detach = filt_fsdetach,
6487 	.f_event = filt_fsevent,
6488 };
6489 
6490 static int
filt_fsattach(struct knote * kn)6491 filt_fsattach(struct knote *kn)
6492 {
6493 
6494 	kn->kn_flags |= EV_CLEAR;
6495 	knlist_add(&fs_knlist, kn, 0);
6496 	return (0);
6497 }
6498 
6499 static void
filt_fsdetach(struct knote * kn)6500 filt_fsdetach(struct knote *kn)
6501 {
6502 
6503 	knlist_remove(&fs_knlist, kn, 0);
6504 }
6505 
6506 static int
filt_fsevent(struct knote * kn,long hint)6507 filt_fsevent(struct knote *kn, long hint)
6508 {
6509 
6510 	kn->kn_fflags |= kn->kn_sfflags & hint;
6511 
6512 	return (kn->kn_fflags != 0);
6513 }
6514 
6515 static int
sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)6516 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6517 {
6518 	struct vfsidctl vc;
6519 	int error;
6520 	struct mount *mp;
6521 
6522 	error = SYSCTL_IN(req, &vc, sizeof(vc));
6523 	if (error)
6524 		return (error);
6525 	if (vc.vc_vers != VFS_CTL_VERS1)
6526 		return (EINVAL);
6527 	mp = vfs_getvfs(&vc.vc_fsid);
6528 	if (mp == NULL)
6529 		return (ENOENT);
6530 	/* ensure that a specific sysctl goes to the right filesystem. */
6531 	if (strcmp(vc.vc_fstypename, "*") != 0 &&
6532 	    strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6533 		vfs_rel(mp);
6534 		return (EINVAL);
6535 	}
6536 	VCTLTOREQ(&vc, req);
6537 	error = VFS_SYSCTL(mp, vc.vc_op, req);
6538 	vfs_rel(mp);
6539 	return (error);
6540 }
6541 
6542 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6543     NULL, 0, sysctl_vfs_ctl, "",
6544     "Sysctl by fsid");
6545 
6546 /*
6547  * Function to initialize a va_filerev field sensibly.
6548  * XXX: Wouldn't a random number make a lot more sense ??
6549  */
6550 u_quad_t
init_va_filerev(void)6551 init_va_filerev(void)
6552 {
6553 	struct bintime bt;
6554 
6555 	getbinuptime(&bt);
6556 	return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6557 }
6558 
6559 static int	filt_vfsread(struct knote *kn, long hint);
6560 static int	filt_vfswrite(struct knote *kn, long hint);
6561 static int	filt_vfsvnode(struct knote *kn, long hint);
6562 static void	filt_vfsdetach(struct knote *kn);
6563 static int	filt_vfsdump(struct proc *p, struct knote *kn,
6564 		    struct kinfo_knote *kin);
6565 
6566 static const struct filterops vfsread_filtops = {
6567 	.f_isfd = 1,
6568 	.f_detach = filt_vfsdetach,
6569 	.f_event = filt_vfsread,
6570 	.f_userdump = filt_vfsdump,
6571 };
6572 static const struct filterops vfswrite_filtops = {
6573 	.f_isfd = 1,
6574 	.f_detach = filt_vfsdetach,
6575 	.f_event = filt_vfswrite,
6576 	.f_userdump = filt_vfsdump,
6577 };
6578 static const struct filterops vfsvnode_filtops = {
6579 	.f_isfd = 1,
6580 	.f_detach = filt_vfsdetach,
6581 	.f_event = filt_vfsvnode,
6582 	.f_userdump = filt_vfsdump,
6583 };
6584 
6585 static void
vfs_knllock(void * arg)6586 vfs_knllock(void *arg)
6587 {
6588 	struct vnode *vp = arg;
6589 
6590 	vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6591 }
6592 
6593 static void
vfs_knlunlock(void * arg)6594 vfs_knlunlock(void *arg)
6595 {
6596 	struct vnode *vp = arg;
6597 
6598 	VOP_UNLOCK(vp);
6599 }
6600 
6601 static void
vfs_knl_assert_lock(void * arg,int what)6602 vfs_knl_assert_lock(void *arg, int what)
6603 {
6604 #ifdef DEBUG_VFS_LOCKS
6605 	struct vnode *vp = arg;
6606 
6607 	if (what == LA_LOCKED)
6608 		ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6609 	else
6610 		ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6611 #endif
6612 }
6613 
6614 int
vfs_kqfilter(struct vop_kqfilter_args * ap)6615 vfs_kqfilter(struct vop_kqfilter_args *ap)
6616 {
6617 	struct vnode *vp = ap->a_vp;
6618 	struct knote *kn = ap->a_kn;
6619 	struct knlist *knl;
6620 
6621 	KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ &&
6622 	    kn->kn_filter != EVFILT_WRITE),
6623 	    ("READ/WRITE filter on a FIFO leaked through"));
6624 	switch (kn->kn_filter) {
6625 	case EVFILT_READ:
6626 		kn->kn_fop = &vfsread_filtops;
6627 		break;
6628 	case EVFILT_WRITE:
6629 		kn->kn_fop = &vfswrite_filtops;
6630 		break;
6631 	case EVFILT_VNODE:
6632 		kn->kn_fop = &vfsvnode_filtops;
6633 		break;
6634 	default:
6635 		return (EINVAL);
6636 	}
6637 
6638 	kn->kn_hook = (caddr_t)vp;
6639 
6640 	v_addpollinfo(vp);
6641 	if (vp->v_pollinfo == NULL)
6642 		return (ENOMEM);
6643 	knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6644 	vhold(vp);
6645 	knlist_add(knl, kn, 0);
6646 
6647 	return (0);
6648 }
6649 
6650 /*
6651  * Detach knote from vnode
6652  */
6653 static void
filt_vfsdetach(struct knote * kn)6654 filt_vfsdetach(struct knote *kn)
6655 {
6656 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6657 
6658 	KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6659 	knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6660 	vdrop(vp);
6661 }
6662 
6663 /*ARGSUSED*/
6664 static int
filt_vfsread(struct knote * kn,long hint)6665 filt_vfsread(struct knote *kn, long hint)
6666 {
6667 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6668 	off_t size;
6669 	int res;
6670 
6671 	/*
6672 	 * filesystem is gone, so set the EOF flag and schedule
6673 	 * the knote for deletion.
6674 	 */
6675 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6676 		VI_LOCK(vp);
6677 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6678 		VI_UNLOCK(vp);
6679 		return (1);
6680 	}
6681 
6682 	if (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0)
6683 		return (0);
6684 
6685 	VI_LOCK(vp);
6686 	kn->kn_data = size - kn->kn_fp->f_offset;
6687 	res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6688 	VI_UNLOCK(vp);
6689 	return (res);
6690 }
6691 
6692 /*ARGSUSED*/
6693 static int
filt_vfswrite(struct knote * kn,long hint)6694 filt_vfswrite(struct knote *kn, long hint)
6695 {
6696 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6697 
6698 	VI_LOCK(vp);
6699 
6700 	/*
6701 	 * filesystem is gone, so set the EOF flag and schedule
6702 	 * the knote for deletion.
6703 	 */
6704 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6705 		kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6706 
6707 	kn->kn_data = 0;
6708 	VI_UNLOCK(vp);
6709 	return (1);
6710 }
6711 
6712 static int
filt_vfsvnode(struct knote * kn,long hint)6713 filt_vfsvnode(struct knote *kn, long hint)
6714 {
6715 	struct vnode *vp = (struct vnode *)kn->kn_hook;
6716 	int res;
6717 
6718 	VI_LOCK(vp);
6719 	if (kn->kn_sfflags & hint)
6720 		kn->kn_fflags |= hint;
6721 	if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6722 		kn->kn_flags |= EV_EOF;
6723 		VI_UNLOCK(vp);
6724 		return (1);
6725 	}
6726 	res = (kn->kn_fflags != 0);
6727 	VI_UNLOCK(vp);
6728 	return (res);
6729 }
6730 
6731 static int
filt_vfsdump(struct proc * p,struct knote * kn,struct kinfo_knote * kin)6732 filt_vfsdump(struct proc *p, struct knote *kn, struct kinfo_knote *kin)
6733 {
6734 	struct vattr va;
6735 	struct vnode *vp;
6736 	char *fullpath, *freepath;
6737 	int error;
6738 
6739 	kin->knt_extdata = KNOTE_EXTDATA_VNODE;
6740 
6741 	vp = kn->kn_fp->f_vnode;
6742 	kin->knt_vnode.knt_vnode_type = vntype_to_kinfo(vp->v_type);
6743 
6744 	va.va_fsid = VNOVAL;
6745 	vn_lock(vp, LK_SHARED | LK_RETRY);
6746 	error = VOP_GETATTR(vp, &va, curthread->td_ucred);
6747 	VOP_UNLOCK(vp);
6748 	if (error != 0)
6749 		return (error);
6750 	kin->knt_vnode.knt_vnode_fsid = va.va_fsid;
6751 	kin->knt_vnode.knt_vnode_fileid = va.va_fileid;
6752 
6753 	freepath = NULL;
6754 	fullpath = "-";
6755 	error = vn_fullpath(vp, &fullpath, &freepath);
6756 	if (error == 0) {
6757 		strlcpy(kin->knt_vnode.knt_vnode_fullpath, fullpath,
6758 		    sizeof(kin->knt_vnode.knt_vnode_fullpath));
6759 	}
6760 	if (freepath != NULL)
6761 		free(freepath, M_TEMP);
6762 
6763 	return (0);
6764 }
6765 
6766 int
vfs_read_dirent(struct vop_readdir_args * ap,struct dirent * dp,off_t off)6767 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6768 {
6769 	int error;
6770 
6771 	if (dp->d_reclen > ap->a_uio->uio_resid)
6772 		return (ENAMETOOLONG);
6773 	error = uiomove(dp, dp->d_reclen, ap->a_uio);
6774 	if (error) {
6775 		if (ap->a_ncookies != NULL) {
6776 			if (ap->a_cookies != NULL)
6777 				free(ap->a_cookies, M_TEMP);
6778 			ap->a_cookies = NULL;
6779 			*ap->a_ncookies = 0;
6780 		}
6781 		return (error);
6782 	}
6783 	if (ap->a_ncookies == NULL)
6784 		return (0);
6785 
6786 	KASSERT(ap->a_cookies,
6787 	    ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6788 
6789 	*ap->a_cookies = realloc(*ap->a_cookies,
6790 	    (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO);
6791 	(*ap->a_cookies)[*ap->a_ncookies] = off;
6792 	*ap->a_ncookies += 1;
6793 	return (0);
6794 }
6795 
6796 /*
6797  * The purpose of this routine is to remove granularity from accmode_t,
6798  * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6799  * VADMIN and VAPPEND.
6800  *
6801  * If it returns 0, the caller is supposed to continue with the usual
6802  * access checks using 'accmode' as modified by this routine.  If it
6803  * returns nonzero value, the caller is supposed to return that value
6804  * as errno.
6805  *
6806  * Note that after this routine runs, accmode may be zero.
6807  */
6808 int
vfs_unixify_accmode(accmode_t * accmode)6809 vfs_unixify_accmode(accmode_t *accmode)
6810 {
6811 	/*
6812 	 * There is no way to specify explicit "deny" rule using
6813 	 * file mode or POSIX.1e ACLs.
6814 	 */
6815 	if (*accmode & VEXPLICIT_DENY) {
6816 		*accmode = 0;
6817 		return (0);
6818 	}
6819 
6820 	/*
6821 	 * None of these can be translated into usual access bits.
6822 	 * Also, the common case for NFSv4 ACLs is to not contain
6823 	 * either of these bits. Caller should check for VWRITE
6824 	 * on the containing directory instead.
6825 	 */
6826 	if (*accmode & (VDELETE_CHILD | VDELETE))
6827 		return (EPERM);
6828 
6829 	if (*accmode & VADMIN_PERMS) {
6830 		*accmode &= ~VADMIN_PERMS;
6831 		*accmode |= VADMIN;
6832 	}
6833 
6834 	/*
6835 	 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6836 	 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6837 	 */
6838 	*accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6839 
6840 	return (0);
6841 }
6842 
6843 /*
6844  * Clear out a doomed vnode (if any) and replace it with a new one as long
6845  * as the fs is not being unmounted. Return the root vnode to the caller.
6846  */
6847 static int __noinline
vfs_cache_root_fallback(struct mount * mp,int flags,struct vnode ** vpp)6848 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6849 {
6850 	struct vnode *vp;
6851 	int error;
6852 
6853 restart:
6854 	if (mp->mnt_rootvnode != NULL) {
6855 		MNT_ILOCK(mp);
6856 		vp = mp->mnt_rootvnode;
6857 		if (vp != NULL) {
6858 			if (!VN_IS_DOOMED(vp)) {
6859 				vrefact(vp);
6860 				MNT_IUNLOCK(mp);
6861 				error = vn_lock(vp, flags);
6862 				if (error == 0) {
6863 					*vpp = vp;
6864 					return (0);
6865 				}
6866 				vrele(vp);
6867 				goto restart;
6868 			}
6869 			/*
6870 			 * Clear the old one.
6871 			 */
6872 			mp->mnt_rootvnode = NULL;
6873 		}
6874 		MNT_IUNLOCK(mp);
6875 		if (vp != NULL) {
6876 			vfs_op_barrier_wait(mp);
6877 			vrele(vp);
6878 		}
6879 	}
6880 	error = VFS_CACHEDROOT(mp, flags, vpp);
6881 	if (error != 0)
6882 		return (error);
6883 	if (mp->mnt_vfs_ops == 0) {
6884 		MNT_ILOCK(mp);
6885 		if (mp->mnt_vfs_ops != 0) {
6886 			MNT_IUNLOCK(mp);
6887 			return (0);
6888 		}
6889 		if (mp->mnt_rootvnode == NULL) {
6890 			vrefact(*vpp);
6891 			mp->mnt_rootvnode = *vpp;
6892 		} else {
6893 			if (mp->mnt_rootvnode != *vpp) {
6894 				if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6895 					panic("%s: mismatch between vnode returned "
6896 					    " by VFS_CACHEDROOT and the one cached "
6897 					    " (%p != %p)",
6898 					    __func__, *vpp, mp->mnt_rootvnode);
6899 				}
6900 			}
6901 		}
6902 		MNT_IUNLOCK(mp);
6903 	}
6904 	return (0);
6905 }
6906 
6907 int
vfs_cache_root(struct mount * mp,int flags,struct vnode ** vpp)6908 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6909 {
6910 	struct mount_pcpu *mpcpu;
6911 	struct vnode *vp;
6912 	int error;
6913 
6914 	if (!vfs_op_thread_enter(mp, mpcpu))
6915 		return (vfs_cache_root_fallback(mp, flags, vpp));
6916 	vp = atomic_load_ptr(&mp->mnt_rootvnode);
6917 	if (vp == NULL || VN_IS_DOOMED(vp)) {
6918 		vfs_op_thread_exit(mp, mpcpu);
6919 		return (vfs_cache_root_fallback(mp, flags, vpp));
6920 	}
6921 	vrefact(vp);
6922 	vfs_op_thread_exit(mp, mpcpu);
6923 	error = vn_lock(vp, flags);
6924 	if (error != 0) {
6925 		vrele(vp);
6926 		return (vfs_cache_root_fallback(mp, flags, vpp));
6927 	}
6928 	*vpp = vp;
6929 	return (0);
6930 }
6931 
6932 struct vnode *
vfs_cache_root_clear(struct mount * mp)6933 vfs_cache_root_clear(struct mount *mp)
6934 {
6935 	struct vnode *vp;
6936 
6937 	/*
6938 	 * ops > 0 guarantees there is nobody who can see this vnode
6939 	 */
6940 	MPASS(mp->mnt_vfs_ops > 0);
6941 	vp = mp->mnt_rootvnode;
6942 	if (vp != NULL)
6943 		vn_seqc_write_begin(vp);
6944 	mp->mnt_rootvnode = NULL;
6945 	return (vp);
6946 }
6947 
6948 void
vfs_cache_root_set(struct mount * mp,struct vnode * vp)6949 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6950 {
6951 
6952 	MPASS(mp->mnt_vfs_ops > 0);
6953 	vrefact(vp);
6954 	mp->mnt_rootvnode = vp;
6955 }
6956 
6957 /*
6958  * These are helper functions for filesystems to traverse all
6959  * their vnodes.  See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6960  *
6961  * This interface replaces MNT_VNODE_FOREACH.
6962  */
6963 
6964 struct vnode *
__mnt_vnode_next_all(struct vnode ** mvp,struct mount * mp)6965 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6966 {
6967 	struct vnode *vp;
6968 
6969 	maybe_yield();
6970 	MNT_ILOCK(mp);
6971 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6972 	for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6973 	    vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6974 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6975 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6976 			continue;
6977 		VI_LOCK(vp);
6978 		if (VN_IS_DOOMED(vp)) {
6979 			VI_UNLOCK(vp);
6980 			continue;
6981 		}
6982 		break;
6983 	}
6984 	if (vp == NULL) {
6985 		__mnt_vnode_markerfree_all(mvp, mp);
6986 		/* MNT_IUNLOCK(mp); -- done in above function */
6987 		mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6988 		return (NULL);
6989 	}
6990 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6991 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6992 	MNT_IUNLOCK(mp);
6993 	return (vp);
6994 }
6995 
6996 struct vnode *
__mnt_vnode_first_all(struct vnode ** mvp,struct mount * mp)6997 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6998 {
6999 	struct vnode *vp;
7000 
7001 	*mvp = vn_alloc_marker(mp);
7002 	MNT_ILOCK(mp);
7003 	MNT_REF(mp);
7004 
7005 	TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
7006 		/* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
7007 		if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
7008 			continue;
7009 		VI_LOCK(vp);
7010 		if (VN_IS_DOOMED(vp)) {
7011 			VI_UNLOCK(vp);
7012 			continue;
7013 		}
7014 		break;
7015 	}
7016 	if (vp == NULL) {
7017 		MNT_REL(mp);
7018 		MNT_IUNLOCK(mp);
7019 		vn_free_marker(*mvp);
7020 		*mvp = NULL;
7021 		return (NULL);
7022 	}
7023 	TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
7024 	MNT_IUNLOCK(mp);
7025 	return (vp);
7026 }
7027 
7028 void
__mnt_vnode_markerfree_all(struct vnode ** mvp,struct mount * mp)7029 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
7030 {
7031 
7032 	if (*mvp == NULL) {
7033 		MNT_IUNLOCK(mp);
7034 		return;
7035 	}
7036 
7037 	mtx_assert(MNT_MTX(mp), MA_OWNED);
7038 
7039 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
7040 	TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
7041 	MNT_REL(mp);
7042 	MNT_IUNLOCK(mp);
7043 	vn_free_marker(*mvp);
7044 	*mvp = NULL;
7045 }
7046 
7047 /*
7048  * These are helper functions for filesystems to traverse their
7049  * lazy vnodes.  See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
7050  */
7051 static void
mnt_vnode_markerfree_lazy(struct vnode ** mvp,struct mount * mp)7052 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
7053 {
7054 
7055 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
7056 
7057 	MNT_ILOCK(mp);
7058 	MNT_REL(mp);
7059 	MNT_IUNLOCK(mp);
7060 	vn_free_marker(*mvp);
7061 	*mvp = NULL;
7062 }
7063 
7064 /*
7065  * Relock the mp mount vnode list lock with the vp vnode interlock in the
7066  * conventional lock order during mnt_vnode_next_lazy iteration.
7067  *
7068  * On entry, the mount vnode list lock is held and the vnode interlock is not.
7069  * The list lock is dropped and reacquired.  On success, both locks are held.
7070  * On failure, the mount vnode list lock is held but the vnode interlock is
7071  * not, and the procedure may have yielded.
7072  */
7073 static bool
mnt_vnode_next_lazy_relock(struct vnode * mvp,struct mount * mp,struct vnode * vp)7074 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
7075     struct vnode *vp)
7076 {
7077 
7078 	VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
7079 	    TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
7080 	    ("%s: bad marker", __func__));
7081 	VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
7082 	    ("%s: inappropriate vnode", __func__));
7083 	ASSERT_VI_UNLOCKED(vp, __func__);
7084 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
7085 
7086 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
7087 	TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
7088 
7089 	/*
7090 	 * Note we may be racing against vdrop which transitioned the hold
7091 	 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
7092 	 * if we are the only user after we get the interlock we will just
7093 	 * vdrop.
7094 	 */
7095 	vhold(vp);
7096 	mtx_unlock(&mp->mnt_listmtx);
7097 	VI_LOCK(vp);
7098 	if (VN_IS_DOOMED(vp)) {
7099 		VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
7100 		goto out_lost;
7101 	}
7102 	VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
7103 	/*
7104 	 * There is nothing to do if we are the last user.
7105 	 */
7106 	if (!refcount_release_if_not_last(&vp->v_holdcnt))
7107 		goto out_lost;
7108 	mtx_lock(&mp->mnt_listmtx);
7109 	return (true);
7110 out_lost:
7111 	vdropl(vp);
7112 	maybe_yield();
7113 	mtx_lock(&mp->mnt_listmtx);
7114 	return (false);
7115 }
7116 
7117 static struct vnode *
mnt_vnode_next_lazy(struct vnode ** mvp,struct mount * mp,mnt_lazy_cb_t * cb,void * cbarg)7118 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
7119     void *cbarg)
7120 {
7121 	struct vnode *vp;
7122 
7123 	mtx_assert(&mp->mnt_listmtx, MA_OWNED);
7124 	KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
7125 restart:
7126 	vp = TAILQ_NEXT(*mvp, v_lazylist);
7127 	while (vp != NULL) {
7128 		if (vp->v_type == VMARKER) {
7129 			vp = TAILQ_NEXT(vp, v_lazylist);
7130 			continue;
7131 		}
7132 		/*
7133 		 * See if we want to process the vnode. Note we may encounter a
7134 		 * long string of vnodes we don't care about and hog the list
7135 		 * as a result. Check for it and requeue the marker.
7136 		 */
7137 		VNPASS(!VN_IS_DOOMED(vp), vp);
7138 		if (!cb(vp, cbarg)) {
7139 			if (!should_yield()) {
7140 				vp = TAILQ_NEXT(vp, v_lazylist);
7141 				continue;
7142 			}
7143 			TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
7144 			    v_lazylist);
7145 			TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
7146 			    v_lazylist);
7147 			mtx_unlock(&mp->mnt_listmtx);
7148 			kern_yield(PRI_USER);
7149 			mtx_lock(&mp->mnt_listmtx);
7150 			goto restart;
7151 		}
7152 		/*
7153 		 * Try-lock because this is the wrong lock order.
7154 		 */
7155 		if (!VI_TRYLOCK(vp) &&
7156 		    !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
7157 			goto restart;
7158 		KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
7159 		KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
7160 		    ("alien vnode on the lazy list %p %p", vp, mp));
7161 		VNPASS(vp->v_mount == mp, vp);
7162 		VNPASS(!VN_IS_DOOMED(vp), vp);
7163 		break;
7164 	}
7165 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
7166 
7167 	/* Check if we are done */
7168 	if (vp == NULL) {
7169 		mtx_unlock(&mp->mnt_listmtx);
7170 		mnt_vnode_markerfree_lazy(mvp, mp);
7171 		return (NULL);
7172 	}
7173 	TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
7174 	mtx_unlock(&mp->mnt_listmtx);
7175 	ASSERT_VI_LOCKED(vp, "lazy iter");
7176 	return (vp);
7177 }
7178 
7179 struct vnode *
__mnt_vnode_next_lazy(struct vnode ** mvp,struct mount * mp,mnt_lazy_cb_t * cb,void * cbarg)7180 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
7181     void *cbarg)
7182 {
7183 
7184 	maybe_yield();
7185 	mtx_lock(&mp->mnt_listmtx);
7186 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
7187 }
7188 
7189 struct vnode *
__mnt_vnode_first_lazy(struct vnode ** mvp,struct mount * mp,mnt_lazy_cb_t * cb,void * cbarg)7190 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
7191     void *cbarg)
7192 {
7193 	struct vnode *vp;
7194 
7195 	if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
7196 		return (NULL);
7197 
7198 	*mvp = vn_alloc_marker(mp);
7199 	MNT_ILOCK(mp);
7200 	MNT_REF(mp);
7201 	MNT_IUNLOCK(mp);
7202 
7203 	mtx_lock(&mp->mnt_listmtx);
7204 	vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
7205 	if (vp == NULL) {
7206 		mtx_unlock(&mp->mnt_listmtx);
7207 		mnt_vnode_markerfree_lazy(mvp, mp);
7208 		return (NULL);
7209 	}
7210 	TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
7211 	return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
7212 }
7213 
7214 void
__mnt_vnode_markerfree_lazy(struct vnode ** mvp,struct mount * mp)7215 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
7216 {
7217 
7218 	if (*mvp == NULL)
7219 		return;
7220 
7221 	mtx_lock(&mp->mnt_listmtx);
7222 	TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
7223 	mtx_unlock(&mp->mnt_listmtx);
7224 	mnt_vnode_markerfree_lazy(mvp, mp);
7225 }
7226 
7227 int
vn_dir_check_exec(struct vnode * vp,struct componentname * cnp)7228 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
7229 {
7230 
7231 	if ((cnp->cn_flags & NOEXECCHECK) != 0) {
7232 		cnp->cn_flags &= ~NOEXECCHECK;
7233 		return (0);
7234 	}
7235 
7236 	return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread));
7237 }
7238 
7239 /*
7240  * Do not use this variant unless you have means other than the hold count
7241  * to prevent the vnode from getting freed.
7242  */
7243 void
vn_seqc_write_begin_locked(struct vnode * vp)7244 vn_seqc_write_begin_locked(struct vnode *vp)
7245 {
7246 
7247 	ASSERT_VI_LOCKED(vp, __func__);
7248 	VNPASS(vp->v_holdcnt > 0, vp);
7249 	VNPASS(vp->v_seqc_users >= 0, vp);
7250 	vp->v_seqc_users++;
7251 	if (vp->v_seqc_users == 1)
7252 		seqc_sleepable_write_begin(&vp->v_seqc);
7253 }
7254 
7255 void
vn_seqc_write_begin(struct vnode * vp)7256 vn_seqc_write_begin(struct vnode *vp)
7257 {
7258 
7259 	VI_LOCK(vp);
7260 	vn_seqc_write_begin_locked(vp);
7261 	VI_UNLOCK(vp);
7262 }
7263 
7264 void
vn_seqc_write_end_locked(struct vnode * vp)7265 vn_seqc_write_end_locked(struct vnode *vp)
7266 {
7267 
7268 	ASSERT_VI_LOCKED(vp, __func__);
7269 	VNPASS(vp->v_seqc_users > 0, vp);
7270 	vp->v_seqc_users--;
7271 	if (vp->v_seqc_users == 0)
7272 		seqc_sleepable_write_end(&vp->v_seqc);
7273 }
7274 
7275 void
vn_seqc_write_end(struct vnode * vp)7276 vn_seqc_write_end(struct vnode *vp)
7277 {
7278 
7279 	VI_LOCK(vp);
7280 	vn_seqc_write_end_locked(vp);
7281 	VI_UNLOCK(vp);
7282 }
7283 
7284 /*
7285  * Special case handling for allocating and freeing vnodes.
7286  *
7287  * The counter remains unchanged on free so that a doomed vnode will
7288  * keep testing as in modify as long as it is accessible with SMR.
7289  */
7290 static void
vn_seqc_init(struct vnode * vp)7291 vn_seqc_init(struct vnode *vp)
7292 {
7293 
7294 	vp->v_seqc = 0;
7295 	vp->v_seqc_users = 0;
7296 }
7297 
7298 static void
vn_seqc_write_end_free(struct vnode * vp)7299 vn_seqc_write_end_free(struct vnode *vp)
7300 {
7301 
7302 	VNPASS(seqc_in_modify(vp->v_seqc), vp);
7303 	VNPASS(vp->v_seqc_users == 1, vp);
7304 }
7305 
7306 void
vn_irflag_set_locked(struct vnode * vp,short toset)7307 vn_irflag_set_locked(struct vnode *vp, short toset)
7308 {
7309 	short flags;
7310 
7311 	ASSERT_VI_LOCKED(vp, __func__);
7312 	flags = vn_irflag_read(vp);
7313 	VNASSERT((flags & toset) == 0, vp,
7314 	    ("%s: some of the passed flags already set (have %d, passed %d)\n",
7315 	    __func__, flags, toset));
7316 	atomic_store_short(&vp->v_irflag, flags | toset);
7317 }
7318 
7319 void
vn_irflag_set(struct vnode * vp,short toset)7320 vn_irflag_set(struct vnode *vp, short toset)
7321 {
7322 
7323 	VI_LOCK(vp);
7324 	vn_irflag_set_locked(vp, toset);
7325 	VI_UNLOCK(vp);
7326 }
7327 
7328 void
vn_irflag_set_cond_locked(struct vnode * vp,short toset)7329 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
7330 {
7331 	short flags;
7332 
7333 	ASSERT_VI_LOCKED(vp, __func__);
7334 	flags = vn_irflag_read(vp);
7335 	atomic_store_short(&vp->v_irflag, flags | toset);
7336 }
7337 
7338 void
vn_irflag_set_cond(struct vnode * vp,short toset)7339 vn_irflag_set_cond(struct vnode *vp, short toset)
7340 {
7341 
7342 	VI_LOCK(vp);
7343 	vn_irflag_set_cond_locked(vp, toset);
7344 	VI_UNLOCK(vp);
7345 }
7346 
7347 void
vn_irflag_unset_locked(struct vnode * vp,short tounset)7348 vn_irflag_unset_locked(struct vnode *vp, short tounset)
7349 {
7350 	short flags;
7351 
7352 	ASSERT_VI_LOCKED(vp, __func__);
7353 	flags = vn_irflag_read(vp);
7354 	VNASSERT((flags & tounset) == tounset, vp,
7355 	    ("%s: some of the passed flags not set (have %d, passed %d)\n",
7356 	    __func__, flags, tounset));
7357 	atomic_store_short(&vp->v_irflag, flags & ~tounset);
7358 }
7359 
7360 void
vn_irflag_unset(struct vnode * vp,short tounset)7361 vn_irflag_unset(struct vnode *vp, short tounset)
7362 {
7363 
7364 	VI_LOCK(vp);
7365 	vn_irflag_unset_locked(vp, tounset);
7366 	VI_UNLOCK(vp);
7367 }
7368 
7369 int
vn_getsize_locked(struct vnode * vp,off_t * size,struct ucred * cred)7370 vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred)
7371 {
7372 	struct vattr vattr;
7373 	int error;
7374 
7375 	ASSERT_VOP_LOCKED(vp, __func__);
7376 	error = VOP_GETATTR(vp, &vattr, cred);
7377 	if (__predict_true(error == 0)) {
7378 		if (vattr.va_size <= OFF_MAX)
7379 			*size = vattr.va_size;
7380 		else
7381 			error = EFBIG;
7382 	}
7383 	return (error);
7384 }
7385 
7386 int
vn_getsize(struct vnode * vp,off_t * size,struct ucred * cred)7387 vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred)
7388 {
7389 	int error;
7390 
7391 	VOP_LOCK(vp, LK_SHARED);
7392 	error = vn_getsize_locked(vp, size, cred);
7393 	VOP_UNLOCK(vp);
7394 	return (error);
7395 }
7396 
7397 #ifdef INVARIANTS
7398 void
vn_set_state_validate(struct vnode * vp,__enum_uint8 (vstate)state)7399 vn_set_state_validate(struct vnode *vp, __enum_uint8(vstate) state)
7400 {
7401 
7402 	switch (vp->v_state) {
7403 	case VSTATE_UNINITIALIZED:
7404 		switch (state) {
7405 		case VSTATE_CONSTRUCTED:
7406 		case VSTATE_DESTROYING:
7407 			return;
7408 		default:
7409 			break;
7410 		}
7411 		break;
7412 	case VSTATE_CONSTRUCTED:
7413 		ASSERT_VOP_ELOCKED(vp, __func__);
7414 		switch (state) {
7415 		case VSTATE_DESTROYING:
7416 			return;
7417 		default:
7418 			break;
7419 		}
7420 		break;
7421 	case VSTATE_DESTROYING:
7422 		ASSERT_VOP_ELOCKED(vp, __func__);
7423 		switch (state) {
7424 		case VSTATE_DEAD:
7425 			return;
7426 		default:
7427 			break;
7428 		}
7429 		break;
7430 	case VSTATE_DEAD:
7431 		switch (state) {
7432 		case VSTATE_UNINITIALIZED:
7433 			return;
7434 		default:
7435 			break;
7436 		}
7437 		break;
7438 	}
7439 
7440 	vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state);
7441 	panic("invalid state transition %d -> %d\n", vp->v_state, state);
7442 }
7443 #endif
7444