FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_subr.c
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 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95
37 */
38
39 /*
40 * External virtual filesystem routines
41 */
42
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45
46 #include "opt_ddb.h"
47 #include "opt_watchdog.h"
48
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/asan.h>
52 #include <sys/bio.h>
53 #include <sys/buf.h>
54 #include <sys/capsicum.h>
55 #include <sys/condvar.h>
56 #include <sys/conf.h>
57 #include <sys/counter.h>
58 #include <sys/dirent.h>
59 #include <sys/event.h>
60 #include <sys/eventhandler.h>
61 #include <sys/extattr.h>
62 #include <sys/file.h>
63 #include <sys/fcntl.h>
64 #include <sys/jail.h>
65 #include <sys/kdb.h>
66 #include <sys/kernel.h>
67 #include <sys/kthread.h>
68 #include <sys/ktr.h>
69 #include <sys/limits.h>
70 #include <sys/lockf.h>
71 #include <sys/malloc.h>
72 #include <sys/mount.h>
73 #include <sys/namei.h>
74 #include <sys/pctrie.h>
75 #include <sys/priv.h>
76 #include <sys/reboot.h>
77 #include <sys/refcount.h>
78 #include <sys/rwlock.h>
79 #include <sys/sched.h>
80 #include <sys/sleepqueue.h>
81 #include <sys/smr.h>
82 #include <sys/smp.h>
83 #include <sys/stat.h>
84 #include <sys/sysctl.h>
85 #include <sys/syslog.h>
86 #include <sys/vmmeter.h>
87 #include <sys/vnode.h>
88 #include <sys/watchdog.h>
89
90 #include <machine/stdarg.h>
91
92 #include <security/mac/mac_framework.h>
93
94 #include <vm/vm.h>
95 #include <vm/vm_object.h>
96 #include <vm/vm_extern.h>
97 #include <vm/pmap.h>
98 #include <vm/vm_map.h>
99 #include <vm/vm_page.h>
100 #include <vm/vm_kern.h>
101 #include <vm/uma.h>
102
103 #if defined(DEBUG_VFS_LOCKS) && (!defined(INVARIANTS) || !defined(WITNESS))
104 #error DEBUG_VFS_LOCKS requires INVARIANTS and WITNESS
105 #endif
106
107 #ifdef DDB
108 #include <ddb/ddb.h>
109 #endif
110
111 static void delmntque(struct vnode *vp);
112 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
113 int slpflag, int slptimeo);
114 static void syncer_shutdown(void *arg, int howto);
115 static int vtryrecycle(struct vnode *vp);
116 static void v_init_counters(struct vnode *);
117 static void vn_seqc_init(struct vnode *);
118 static void vn_seqc_write_end_free(struct vnode *vp);
119 static void vgonel(struct vnode *);
120 static bool vhold_recycle_free(struct vnode *);
121 static void vdropl_recycle(struct vnode *vp);
122 static void vdrop_recycle(struct vnode *vp);
123 static void vfs_knllock(void *arg);
124 static void vfs_knlunlock(void *arg);
125 static void vfs_knl_assert_lock(void *arg, int what);
126 static void destroy_vpollinfo(struct vpollinfo *vi);
127 static int v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
128 daddr_t startlbn, daddr_t endlbn);
129 static void vnlru_recalc(void);
130
131 /*
132 * Number of vnodes in existence. Increased whenever getnewvnode()
133 * allocates a new vnode, decreased in vdropl() for VIRF_DOOMED vnode.
134 */
135 static u_long __exclusive_cache_line numvnodes;
136
137 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
138 "Number of vnodes in existence");
139
140 static counter_u64_t vnodes_created;
141 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
142 "Number of vnodes created by getnewvnode");
143
144 /*
145 * Conversion tables for conversion from vnode types to inode formats
146 * and back.
147 */
148 enum vtype iftovt_tab[16] = {
149 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
150 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VNON
151 };
152 int vttoif_tab[10] = {
153 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
154 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
155 };
156
157 /*
158 * List of allocates vnodes in the system.
159 */
160 static TAILQ_HEAD(freelst, vnode) vnode_list;
161 static struct vnode *vnode_list_free_marker;
162 static struct vnode *vnode_list_reclaim_marker;
163
164 /*
165 * "Free" vnode target. Free vnodes are rarely completely free, but are
166 * just ones that are cheap to recycle. Usually they are for files which
167 * have been stat'd but not read; these usually have inode and namecache
168 * data attached to them. This target is the preferred minimum size of a
169 * sub-cache consisting mostly of such files. The system balances the size
170 * of this sub-cache with its complement to try to prevent either from
171 * thrashing while the other is relatively inactive. The targets express
172 * a preference for the best balance.
173 *
174 * "Above" this target there are 2 further targets (watermarks) related
175 * to recyling of free vnodes. In the best-operating case, the cache is
176 * exactly full, the free list has size between vlowat and vhiwat above the
177 * free target, and recycling from it and normal use maintains this state.
178 * Sometimes the free list is below vlowat or even empty, but this state
179 * is even better for immediate use provided the cache is not full.
180 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
181 * ones) to reach one of these states. The watermarks are currently hard-
182 * coded as 4% and 9% of the available space higher. These and the default
183 * of 25% for wantfreevnodes are too large if the memory size is large.
184 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
185 * whenever vnlru_proc() becomes active.
186 */
187 static long wantfreevnodes;
188 static long __exclusive_cache_line freevnodes;
189 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
190 &freevnodes, 0, "Number of \"free\" vnodes");
191 static long freevnodes_old;
192
193 static counter_u64_t recycles_count;
194 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
195 "Number of vnodes recycled to meet vnode cache targets");
196
197 static counter_u64_t recycles_free_count;
198 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles_free, CTLFLAG_RD, &recycles_free_count,
199 "Number of free vnodes recycled to meet vnode cache targets");
200
201 static counter_u64_t deferred_inact;
202 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
203 "Number of times inactive processing was deferred");
204
205 /* To keep more than one thread at a time from running vfs_getnewfsid */
206 static struct mtx mntid_mtx;
207
208 /*
209 * Lock for any access to the following:
210 * vnode_list
211 * numvnodes
212 * freevnodes
213 */
214 static struct mtx __exclusive_cache_line vnode_list_mtx;
215
216 /* Publicly exported FS */
217 struct nfs_public nfs_pub;
218
219 static uma_zone_t buf_trie_zone;
220 static smr_t buf_trie_smr;
221
222 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
223 static uma_zone_t vnode_zone;
224 MALLOC_DEFINE(M_VNODEPOLL, "VN POLL", "vnode poll");
225
226 __read_frequently smr_t vfs_smr;
227
228 /*
229 * The workitem queue.
230 *
231 * It is useful to delay writes of file data and filesystem metadata
232 * for tens of seconds so that quickly created and deleted files need
233 * not waste disk bandwidth being created and removed. To realize this,
234 * we append vnodes to a "workitem" queue. When running with a soft
235 * updates implementation, most pending metadata dependencies should
236 * not wait for more than a few seconds. Thus, mounted on block devices
237 * are delayed only about a half the time that file data is delayed.
238 * Similarly, directory updates are more critical, so are only delayed
239 * about a third the time that file data is delayed. Thus, there are
240 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
241 * one each second (driven off the filesystem syncer process). The
242 * syncer_delayno variable indicates the next queue that is to be processed.
243 * Items that need to be processed soon are placed in this queue:
244 *
245 * syncer_workitem_pending[syncer_delayno]
246 *
247 * A delay of fifteen seconds is done by placing the request fifteen
248 * entries later in the queue:
249 *
250 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
251 *
252 */
253 static int syncer_delayno;
254 static long syncer_mask;
255 LIST_HEAD(synclist, bufobj);
256 static struct synclist *syncer_workitem_pending;
257 /*
258 * The sync_mtx protects:
259 * bo->bo_synclist
260 * sync_vnode_count
261 * syncer_delayno
262 * syncer_state
263 * syncer_workitem_pending
264 * syncer_worklist_len
265 * rushjob
266 */
267 static struct mtx sync_mtx;
268 static struct cv sync_wakeup;
269
270 #define SYNCER_MAXDELAY 32
271 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */
272 static int syncdelay = 30; /* max time to delay syncing data */
273 static int filedelay = 30; /* time to delay syncing files */
274 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
275 "Time to delay syncing files (in seconds)");
276 static int dirdelay = 29; /* time to delay syncing directories */
277 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
278 "Time to delay syncing directories (in seconds)");
279 static int metadelay = 28; /* time to delay syncing metadata */
280 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
281 "Time to delay syncing metadata (in seconds)");
282 static int rushjob; /* number of slots to run ASAP */
283 static int stat_rush_requests; /* number of times I/O speeded up */
284 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
285 "Number of times I/O speeded up (rush requests)");
286
287 #define VDBATCH_SIZE 8
288 struct vdbatch {
289 u_int index;
290 long freevnodes;
291 struct mtx lock;
292 struct vnode *tab[VDBATCH_SIZE];
293 };
294 DPCPU_DEFINE_STATIC(struct vdbatch, vd);
295
296 static void vdbatch_dequeue(struct vnode *vp);
297
298 /*
299 * When shutting down the syncer, run it at four times normal speed.
300 */
301 #define SYNCER_SHUTDOWN_SPEEDUP 4
302 static int sync_vnode_count;
303 static int syncer_worklist_len;
304 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
305 syncer_state;
306
307 /* Target for maximum number of vnodes. */
308 u_long desiredvnodes;
309 static u_long gapvnodes; /* gap between wanted and desired */
310 static u_long vhiwat; /* enough extras after expansion */
311 static u_long vlowat; /* minimal extras before expansion */
312 static u_long vstir; /* nonzero to stir non-free vnodes */
313 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */
314
315 static u_long vnlru_read_freevnodes(void);
316
317 /*
318 * Note that no attempt is made to sanitize these parameters.
319 */
320 static int
321 sysctl_maxvnodes(SYSCTL_HANDLER_ARGS)
322 {
323 u_long val;
324 int error;
325
326 val = desiredvnodes;
327 error = sysctl_handle_long(oidp, &val, 0, req);
328 if (error != 0 || req->newptr == NULL)
329 return (error);
330
331 if (val == desiredvnodes)
332 return (0);
333 mtx_lock(&vnode_list_mtx);
334 desiredvnodes = val;
335 wantfreevnodes = desiredvnodes / 4;
336 vnlru_recalc();
337 mtx_unlock(&vnode_list_mtx);
338 /*
339 * XXX There is no protection against multiple threads changing
340 * desiredvnodes at the same time. Locking above only helps vnlru and
341 * getnewvnode.
342 */
343 vfs_hash_changesize(desiredvnodes);
344 cache_changesize(desiredvnodes);
345 return (0);
346 }
347
348 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
349 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_maxvnodes,
350 "LU", "Target for maximum number of vnodes");
351
352 static int
353 sysctl_wantfreevnodes(SYSCTL_HANDLER_ARGS)
354 {
355 u_long val;
356 int error;
357
358 val = wantfreevnodes;
359 error = sysctl_handle_long(oidp, &val, 0, req);
360 if (error != 0 || req->newptr == NULL)
361 return (error);
362
363 if (val == wantfreevnodes)
364 return (0);
365 mtx_lock(&vnode_list_mtx);
366 wantfreevnodes = val;
367 vnlru_recalc();
368 mtx_unlock(&vnode_list_mtx);
369 return (0);
370 }
371
372 SYSCTL_PROC(_vfs, OID_AUTO, wantfreevnodes,
373 CTLTYPE_ULONG | CTLFLAG_MPSAFE | CTLFLAG_RW, NULL, 0, sysctl_wantfreevnodes,
374 "LU", "Target for minimum number of \"free\" vnodes");
375
376 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
377 &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
378 static int vnlru_nowhere;
379 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW | CTLFLAG_STATS,
380 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
381
382 static int
383 sysctl_try_reclaim_vnode(SYSCTL_HANDLER_ARGS)
384 {
385 struct vnode *vp;
386 struct nameidata nd;
387 char *buf;
388 unsigned long ndflags;
389 int error;
390
391 if (req->newptr == NULL)
392 return (EINVAL);
393 if (req->newlen >= PATH_MAX)
394 return (E2BIG);
395
396 buf = malloc(PATH_MAX, M_TEMP, M_WAITOK);
397 error = SYSCTL_IN(req, buf, req->newlen);
398 if (error != 0)
399 goto out;
400
401 buf[req->newlen] = '\0';
402
403 ndflags = LOCKLEAF | NOFOLLOW | AUDITVNODE1;
404 NDINIT(&nd, LOOKUP, ndflags, UIO_SYSSPACE, buf);
405 if ((error = namei(&nd)) != 0)
406 goto out;
407 vp = nd.ni_vp;
408
409 if (VN_IS_DOOMED(vp)) {
410 /*
411 * This vnode is being recycled. Return != 0 to let the caller
412 * know that the sysctl had no effect. Return EAGAIN because a
413 * subsequent call will likely succeed (since namei will create
414 * a new vnode if necessary)
415 */
416 error = EAGAIN;
417 goto putvnode;
418 }
419
420 counter_u64_add(recycles_count, 1);
421 vgone(vp);
422 putvnode:
423 vput(vp);
424 NDFREE_PNBUF(&nd);
425 out:
426 free(buf, M_TEMP);
427 return (error);
428 }
429
430 static int
431 sysctl_ftry_reclaim_vnode(SYSCTL_HANDLER_ARGS)
432 {
433 struct thread *td = curthread;
434 struct vnode *vp;
435 struct file *fp;
436 int error;
437 int fd;
438
439 if (req->newptr == NULL)
440 return (EBADF);
441
442 error = sysctl_handle_int(oidp, &fd, 0, req);
443 if (error != 0)
444 return (error);
445 error = getvnode(curthread, fd, &cap_fcntl_rights, &fp);
446 if (error != 0)
447 return (error);
448 vp = fp->f_vnode;
449
450 error = vn_lock(vp, LK_EXCLUSIVE);
451 if (error != 0)
452 goto drop;
453
454 counter_u64_add(recycles_count, 1);
455 vgone(vp);
456 VOP_UNLOCK(vp);
457 drop:
458 fdrop(fp, td);
459 return (error);
460 }
461
462 SYSCTL_PROC(_debug, OID_AUTO, try_reclaim_vnode,
463 CTLTYPE_STRING | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
464 sysctl_try_reclaim_vnode, "A", "Try to reclaim a vnode by its pathname");
465 SYSCTL_PROC(_debug, OID_AUTO, ftry_reclaim_vnode,
466 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_WR, NULL, 0,
467 sysctl_ftry_reclaim_vnode, "I",
468 "Try to reclaim a vnode by its file descriptor");
469
470 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
471 #define vnsz2log 8
472 #ifndef DEBUG_LOCKS
473 _Static_assert(sizeof(struct vnode) >= 1UL << vnsz2log &&
474 sizeof(struct vnode) < 1UL << (vnsz2log + 1),
475 "vnsz2log needs to be updated");
476 #endif
477
478 /*
479 * Support for the bufobj clean & dirty pctrie.
480 */
481 static void *
482 buf_trie_alloc(struct pctrie *ptree)
483 {
484 return (uma_zalloc_smr(buf_trie_zone, M_NOWAIT));
485 }
486
487 static void
488 buf_trie_free(struct pctrie *ptree, void *node)
489 {
490 uma_zfree_smr(buf_trie_zone, node);
491 }
492 PCTRIE_DEFINE_SMR(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free,
493 buf_trie_smr);
494
495 /*
496 * Initialize the vnode management data structures.
497 *
498 * Reevaluate the following cap on the number of vnodes after the physical
499 * memory size exceeds 512GB. In the limit, as the physical memory size
500 * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
501 */
502 #ifndef MAXVNODES_MAX
503 #define MAXVNODES_MAX (512UL * 1024 * 1024 / 64) /* 8M */
504 #endif
505
506 static MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
507
508 static struct vnode *
509 vn_alloc_marker(struct mount *mp)
510 {
511 struct vnode *vp;
512
513 vp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
514 vp->v_type = VMARKER;
515 vp->v_mount = mp;
516
517 return (vp);
518 }
519
520 static void
521 vn_free_marker(struct vnode *vp)
522 {
523
524 MPASS(vp->v_type == VMARKER);
525 free(vp, M_VNODE_MARKER);
526 }
527
528 #ifdef KASAN
529 static int
530 vnode_ctor(void *mem, int size, void *arg __unused, int flags __unused)
531 {
532 kasan_mark(mem, size, roundup2(size, UMA_ALIGN_PTR + 1), 0);
533 return (0);
534 }
535
536 static void
537 vnode_dtor(void *mem, int size, void *arg __unused)
538 {
539 size_t end1, end2, off1, off2;
540
541 _Static_assert(offsetof(struct vnode, v_vnodelist) <
542 offsetof(struct vnode, v_dbatchcpu),
543 "KASAN marks require updating");
544
545 off1 = offsetof(struct vnode, v_vnodelist);
546 off2 = offsetof(struct vnode, v_dbatchcpu);
547 end1 = off1 + sizeof(((struct vnode *)NULL)->v_vnodelist);
548 end2 = off2 + sizeof(((struct vnode *)NULL)->v_dbatchcpu);
549
550 /*
551 * Access to the v_vnodelist and v_dbatchcpu fields are permitted even
552 * after the vnode has been freed. Try to get some KASAN coverage by
553 * marking everything except those two fields as invalid. Because
554 * KASAN's tracking is not byte-granular, any preceding fields sharing
555 * the same 8-byte aligned word must also be marked valid.
556 */
557
558 /* Handle the area from the start until v_vnodelist... */
559 off1 = rounddown2(off1, KASAN_SHADOW_SCALE);
560 kasan_mark(mem, off1, off1, KASAN_UMA_FREED);
561
562 /* ... then the area between v_vnodelist and v_dbatchcpu ... */
563 off1 = roundup2(end1, KASAN_SHADOW_SCALE);
564 off2 = rounddown2(off2, KASAN_SHADOW_SCALE);
565 if (off2 > off1)
566 kasan_mark((void *)((char *)mem + off1), off2 - off1,
567 off2 - off1, KASAN_UMA_FREED);
568
569 /* ... and finally the area from v_dbatchcpu to the end. */
570 off2 = roundup2(end2, KASAN_SHADOW_SCALE);
571 kasan_mark((void *)((char *)mem + off2), size - off2, size - off2,
572 KASAN_UMA_FREED);
573 }
574 #endif /* KASAN */
575
576 /*
577 * Initialize a vnode as it first enters the zone.
578 */
579 static int
580 vnode_init(void *mem, int size, int flags)
581 {
582 struct vnode *vp;
583
584 vp = mem;
585 bzero(vp, size);
586 /*
587 * Setup locks.
588 */
589 vp->v_vnlock = &vp->v_lock;
590 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
591 /*
592 * By default, don't allow shared locks unless filesystems opt-in.
593 */
594 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
595 LK_NOSHARE | LK_IS_VNODE);
596 /*
597 * Initialize bufobj.
598 */
599 bufobj_init(&vp->v_bufobj, vp);
600 /*
601 * Initialize namecache.
602 */
603 cache_vnode_init(vp);
604 /*
605 * Initialize rangelocks.
606 */
607 rangelock_init(&vp->v_rl);
608
609 vp->v_dbatchcpu = NOCPU;
610
611 vp->v_state = VSTATE_DEAD;
612
613 /*
614 * Check vhold_recycle_free for an explanation.
615 */
616 vp->v_holdcnt = VHOLD_NO_SMR;
617 vp->v_type = VNON;
618 mtx_lock(&vnode_list_mtx);
619 TAILQ_INSERT_BEFORE(vnode_list_free_marker, vp, v_vnodelist);
620 mtx_unlock(&vnode_list_mtx);
621 return (0);
622 }
623
624 /*
625 * Free a vnode when it is cleared from the zone.
626 */
627 static void
628 vnode_fini(void *mem, int size)
629 {
630 struct vnode *vp;
631 struct bufobj *bo;
632
633 vp = mem;
634 vdbatch_dequeue(vp);
635 mtx_lock(&vnode_list_mtx);
636 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
637 mtx_unlock(&vnode_list_mtx);
638 rangelock_destroy(&vp->v_rl);
639 lockdestroy(vp->v_vnlock);
640 mtx_destroy(&vp->v_interlock);
641 bo = &vp->v_bufobj;
642 rw_destroy(BO_LOCKPTR(bo));
643
644 kasan_mark(mem, size, size, 0);
645 }
646
647 /*
648 * Provide the size of NFS nclnode and NFS fh for calculation of the
649 * vnode memory consumption. The size is specified directly to
650 * eliminate dependency on NFS-private header.
651 *
652 * Other filesystems may use bigger or smaller (like UFS and ZFS)
653 * private inode data, but the NFS-based estimation is ample enough.
654 * Still, we care about differences in the size between 64- and 32-bit
655 * platforms.
656 *
657 * Namecache structure size is heuristically
658 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
659 */
660 #ifdef _LP64
661 #define NFS_NCLNODE_SZ (528 + 64)
662 #define NC_SZ 148
663 #else
664 #define NFS_NCLNODE_SZ (360 + 32)
665 #define NC_SZ 92
666 #endif
667
668 static void
669 vntblinit(void *dummy __unused)
670 {
671 struct vdbatch *vd;
672 uma_ctor ctor;
673 uma_dtor dtor;
674 int cpu, physvnodes, virtvnodes;
675
676 /*
677 * Desiredvnodes is a function of the physical memory size and the
678 * kernel's heap size. Generally speaking, it scales with the
679 * physical memory size. The ratio of desiredvnodes to the physical
680 * memory size is 1:16 until desiredvnodes exceeds 98,304.
681 * Thereafter, the
682 * marginal ratio of desiredvnodes to the physical memory size is
683 * 1:64. However, desiredvnodes is limited by the kernel's heap
684 * size. The memory required by desiredvnodes vnodes and vm objects
685 * must not exceed 1/10th of the kernel's heap size.
686 */
687 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
688 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
689 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
690 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
691 desiredvnodes = min(physvnodes, virtvnodes);
692 if (desiredvnodes > MAXVNODES_MAX) {
693 if (bootverbose)
694 printf("Reducing kern.maxvnodes %lu -> %lu\n",
695 desiredvnodes, MAXVNODES_MAX);
696 desiredvnodes = MAXVNODES_MAX;
697 }
698 wantfreevnodes = desiredvnodes / 4;
699 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
700 TAILQ_INIT(&vnode_list);
701 mtx_init(&vnode_list_mtx, "vnode_list", NULL, MTX_DEF);
702 /*
703 * The lock is taken to appease WITNESS.
704 */
705 mtx_lock(&vnode_list_mtx);
706 vnlru_recalc();
707 mtx_unlock(&vnode_list_mtx);
708 vnode_list_free_marker = vn_alloc_marker(NULL);
709 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_free_marker, v_vnodelist);
710 vnode_list_reclaim_marker = vn_alloc_marker(NULL);
711 TAILQ_INSERT_HEAD(&vnode_list, vnode_list_reclaim_marker, v_vnodelist);
712
713 #ifdef KASAN
714 ctor = vnode_ctor;
715 dtor = vnode_dtor;
716 #else
717 ctor = NULL;
718 dtor = NULL;
719 #endif
720 vnode_zone = uma_zcreate("VNODE", sizeof(struct vnode), ctor, dtor,
721 vnode_init, vnode_fini, UMA_ALIGN_PTR, UMA_ZONE_NOKASAN);
722 uma_zone_set_smr(vnode_zone, vfs_smr);
723
724 /*
725 * Preallocate enough nodes to support one-per buf so that
726 * we can not fail an insert. reassignbuf() callers can not
727 * tolerate the insertion failure.
728 */
729 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
730 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
731 UMA_ZONE_NOFREE | UMA_ZONE_SMR);
732 buf_trie_smr = uma_zone_get_smr(buf_trie_zone);
733 uma_prealloc(buf_trie_zone, nbuf);
734
735 vnodes_created = counter_u64_alloc(M_WAITOK);
736 recycles_count = counter_u64_alloc(M_WAITOK);
737 recycles_free_count = counter_u64_alloc(M_WAITOK);
738 deferred_inact = counter_u64_alloc(M_WAITOK);
739
740 /*
741 * Initialize the filesystem syncer.
742 */
743 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
744 &syncer_mask);
745 syncer_maxdelay = syncer_mask + 1;
746 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
747 cv_init(&sync_wakeup, "syncer");
748
749 CPU_FOREACH(cpu) {
750 vd = DPCPU_ID_PTR((cpu), vd);
751 bzero(vd, sizeof(*vd));
752 mtx_init(&vd->lock, "vdbatch", NULL, MTX_DEF);
753 }
754 }
755 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
756
757 /*
758 * Mark a mount point as busy. Used to synchronize access and to delay
759 * unmounting. Eventually, mountlist_mtx is not released on failure.
760 *
761 * vfs_busy() is a custom lock, it can block the caller.
762 * vfs_busy() only sleeps if the unmount is active on the mount point.
763 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
764 * vnode belonging to mp.
765 *
766 * Lookup uses vfs_busy() to traverse mount points.
767 * root fs var fs
768 * / vnode lock A / vnode lock (/var) D
769 * /var vnode lock B /log vnode lock(/var/log) E
770 * vfs_busy lock C vfs_busy lock F
771 *
772 * Within each file system, the lock order is C->A->B and F->D->E.
773 *
774 * When traversing across mounts, the system follows that lock order:
775 *
776 * C->A->B
777 * |
778 * +->F->D->E
779 *
780 * The lookup() process for namei("/var") illustrates the process:
781 * 1. VOP_LOOKUP() obtains B while A is held
782 * 2. vfs_busy() obtains a shared lock on F while A and B are held
783 * 3. vput() releases lock on B
784 * 4. vput() releases lock on A
785 * 5. VFS_ROOT() obtains lock on D while shared lock on F is held
786 * 6. vfs_unbusy() releases shared lock on F
787 * 7. vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
788 * Attempt to lock A (instead of vp_crossmp) while D is held would
789 * violate the global order, causing deadlocks.
790 *
791 * dounmount() locks B while F is drained. Note that for stacked
792 * filesystems, D and B in the example above may be the same lock,
793 * which introdues potential lock order reversal deadlock between
794 * dounmount() and step 5 above. These filesystems may avoid the LOR
795 * by setting VV_CROSSLOCK on the covered vnode so that lock B will
796 * remain held until after step 5.
797 */
798 int
799 vfs_busy(struct mount *mp, int flags)
800 {
801 struct mount_pcpu *mpcpu;
802
803 MPASS((flags & ~MBF_MASK) == 0);
804 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
805
806 if (vfs_op_thread_enter(mp, mpcpu)) {
807 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
808 MPASS((mp->mnt_kern_flag & MNTK_UNMOUNT) == 0);
809 MPASS((mp->mnt_kern_flag & MNTK_REFEXPIRE) == 0);
810 vfs_mp_count_add_pcpu(mpcpu, ref, 1);
811 vfs_mp_count_add_pcpu(mpcpu, lockref, 1);
812 vfs_op_thread_exit(mp, mpcpu);
813 if (flags & MBF_MNTLSTLOCK)
814 mtx_unlock(&mountlist_mtx);
815 return (0);
816 }
817
818 MNT_ILOCK(mp);
819 vfs_assert_mount_counters(mp);
820 MNT_REF(mp);
821 /*
822 * If mount point is currently being unmounted, sleep until the
823 * mount point fate is decided. If thread doing the unmounting fails,
824 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
825 * that this mount point has survived the unmount attempt and vfs_busy
826 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE
827 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
828 * about to be really destroyed. vfs_busy needs to release its
829 * reference on the mount point in this case and return with ENOENT,
830 * telling the caller the mount it tried to busy is no longer valid.
831 */
832 while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
833 KASSERT(TAILQ_EMPTY(&mp->mnt_uppers),
834 ("%s: non-empty upper mount list with pending unmount",
835 __func__));
836 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
837 MNT_REL(mp);
838 MNT_IUNLOCK(mp);
839 CTR1(KTR_VFS, "%s: failed busying before sleeping",
840 __func__);
841 return (ENOENT);
842 }
843 if (flags & MBF_MNTLSTLOCK)
844 mtx_unlock(&mountlist_mtx);
845 mp->mnt_kern_flag |= MNTK_MWAIT;
846 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
847 if (flags & MBF_MNTLSTLOCK)
848 mtx_lock(&mountlist_mtx);
849 MNT_ILOCK(mp);
850 }
851 if (flags & MBF_MNTLSTLOCK)
852 mtx_unlock(&mountlist_mtx);
853 mp->mnt_lockref++;
854 MNT_IUNLOCK(mp);
855 return (0);
856 }
857
858 /*
859 * Free a busy filesystem.
860 */
861 void
862 vfs_unbusy(struct mount *mp)
863 {
864 struct mount_pcpu *mpcpu;
865 int c;
866
867 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
868
869 if (vfs_op_thread_enter(mp, mpcpu)) {
870 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
871 vfs_mp_count_sub_pcpu(mpcpu, lockref, 1);
872 vfs_mp_count_sub_pcpu(mpcpu, ref, 1);
873 vfs_op_thread_exit(mp, mpcpu);
874 return;
875 }
876
877 MNT_ILOCK(mp);
878 vfs_assert_mount_counters(mp);
879 MNT_REL(mp);
880 c = --mp->mnt_lockref;
881 if (mp->mnt_vfs_ops == 0) {
882 MPASS((mp->mnt_kern_flag & MNTK_DRAINING) == 0);
883 MNT_IUNLOCK(mp);
884 return;
885 }
886 if (c < 0)
887 vfs_dump_mount_counters(mp);
888 if (c == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
889 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
890 CTR1(KTR_VFS, "%s: waking up waiters", __func__);
891 mp->mnt_kern_flag &= ~MNTK_DRAINING;
892 wakeup(&mp->mnt_lockref);
893 }
894 MNT_IUNLOCK(mp);
895 }
896
897 /*
898 * Lookup a mount point by filesystem identifier.
899 */
900 struct mount *
901 vfs_getvfs(fsid_t *fsid)
902 {
903 struct mount *mp;
904
905 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
906 mtx_lock(&mountlist_mtx);
907 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
908 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
909 vfs_ref(mp);
910 mtx_unlock(&mountlist_mtx);
911 return (mp);
912 }
913 }
914 mtx_unlock(&mountlist_mtx);
915 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
916 return ((struct mount *) 0);
917 }
918
919 /*
920 * Lookup a mount point by filesystem identifier, busying it before
921 * returning.
922 *
923 * To avoid congestion on mountlist_mtx, implement simple direct-mapped
924 * cache for popular filesystem identifiers. The cache is lockess, using
925 * the fact that struct mount's are never freed. In worst case we may
926 * get pointer to unmounted or even different filesystem, so we have to
927 * check what we got, and go slow way if so.
928 */
929 struct mount *
930 vfs_busyfs(fsid_t *fsid)
931 {
932 #define FSID_CACHE_SIZE 256
933 typedef struct mount * volatile vmp_t;
934 static vmp_t cache[FSID_CACHE_SIZE];
935 struct mount *mp;
936 int error;
937 uint32_t hash;
938
939 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
940 hash = fsid->val[0] ^ fsid->val[1];
941 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
942 mp = cache[hash];
943 if (mp == NULL || fsidcmp(&mp->mnt_stat.f_fsid, fsid) != 0)
944 goto slow;
945 if (vfs_busy(mp, 0) != 0) {
946 cache[hash] = NULL;
947 goto slow;
948 }
949 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0)
950 return (mp);
951 else
952 vfs_unbusy(mp);
953
954 slow:
955 mtx_lock(&mountlist_mtx);
956 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
957 if (fsidcmp(&mp->mnt_stat.f_fsid, fsid) == 0) {
958 error = vfs_busy(mp, MBF_MNTLSTLOCK);
959 if (error) {
960 cache[hash] = NULL;
961 mtx_unlock(&mountlist_mtx);
962 return (NULL);
963 }
964 cache[hash] = mp;
965 return (mp);
966 }
967 }
968 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
969 mtx_unlock(&mountlist_mtx);
970 return ((struct mount *) 0);
971 }
972
973 /*
974 * Check if a user can access privileged mount options.
975 */
976 int
977 vfs_suser(struct mount *mp, struct thread *td)
978 {
979 int error;
980
981 if (jailed(td->td_ucred)) {
982 /*
983 * If the jail of the calling thread lacks permission for
984 * this type of file system, deny immediately.
985 */
986 if (!prison_allow(td->td_ucred, mp->mnt_vfc->vfc_prison_flag))
987 return (EPERM);
988
989 /*
990 * If the file system was mounted outside the jail of the
991 * calling thread, deny immediately.
992 */
993 if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
994 return (EPERM);
995 }
996
997 /*
998 * If file system supports delegated administration, we don't check
999 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
1000 * by the file system itself.
1001 * If this is not the user that did original mount, we check for
1002 * the PRIV_VFS_MOUNT_OWNER privilege.
1003 */
1004 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
1005 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
1006 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
1007 return (error);
1008 }
1009 return (0);
1010 }
1011
1012 /*
1013 * Get a new unique fsid. Try to make its val[0] unique, since this value
1014 * will be used to create fake device numbers for stat(). Also try (but
1015 * not so hard) make its val[0] unique mod 2^16, since some emulators only
1016 * support 16-bit device numbers. We end up with unique val[0]'s for the
1017 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
1018 *
1019 * Keep in mind that several mounts may be running in parallel. Starting
1020 * the search one past where the previous search terminated is both a
1021 * micro-optimization and a defense against returning the same fsid to
1022 * different mounts.
1023 */
1024 void
1025 vfs_getnewfsid(struct mount *mp)
1026 {
1027 static uint16_t mntid_base;
1028 struct mount *nmp;
1029 fsid_t tfsid;
1030 int mtype;
1031
1032 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
1033 mtx_lock(&mntid_mtx);
1034 mtype = mp->mnt_vfc->vfc_typenum;
1035 tfsid.val[1] = mtype;
1036 mtype = (mtype & 0xFF) << 24;
1037 for (;;) {
1038 tfsid.val[0] = makedev(255,
1039 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
1040 mntid_base++;
1041 if ((nmp = vfs_getvfs(&tfsid)) == NULL)
1042 break;
1043 vfs_rel(nmp);
1044 }
1045 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
1046 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
1047 mtx_unlock(&mntid_mtx);
1048 }
1049
1050 /*
1051 * Knob to control the precision of file timestamps:
1052 *
1053 * 0 = seconds only; nanoseconds zeroed.
1054 * 1 = seconds and nanoseconds, accurate within 1/HZ.
1055 * 2 = seconds and nanoseconds, truncated to microseconds.
1056 * >=3 = seconds and nanoseconds, maximum precision.
1057 */
1058 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
1059
1060 static int timestamp_precision = TSP_USEC;
1061 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
1062 ×tamp_precision, 0, "File timestamp precision (0: seconds, "
1063 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
1064 "3+: sec + ns (max. precision))");
1065
1066 /*
1067 * Get a current timestamp.
1068 */
1069 void
1070 vfs_timestamp(struct timespec *tsp)
1071 {
1072 struct timeval tv;
1073
1074 switch (timestamp_precision) {
1075 case TSP_SEC:
1076 tsp->tv_sec = time_second;
1077 tsp->tv_nsec = 0;
1078 break;
1079 case TSP_HZ:
1080 getnanotime(tsp);
1081 break;
1082 case TSP_USEC:
1083 microtime(&tv);
1084 TIMEVAL_TO_TIMESPEC(&tv, tsp);
1085 break;
1086 case TSP_NSEC:
1087 default:
1088 nanotime(tsp);
1089 break;
1090 }
1091 }
1092
1093 /*
1094 * Set vnode attributes to VNOVAL
1095 */
1096 void
1097 vattr_null(struct vattr *vap)
1098 {
1099
1100 vap->va_type = VNON;
1101 vap->va_size = VNOVAL;
1102 vap->va_bytes = VNOVAL;
1103 vap->va_mode = VNOVAL;
1104 vap->va_nlink = VNOVAL;
1105 vap->va_uid = VNOVAL;
1106 vap->va_gid = VNOVAL;
1107 vap->va_fsid = VNOVAL;
1108 vap->va_fileid = VNOVAL;
1109 vap->va_blocksize = VNOVAL;
1110 vap->va_rdev = VNOVAL;
1111 vap->va_atime.tv_sec = VNOVAL;
1112 vap->va_atime.tv_nsec = VNOVAL;
1113 vap->va_mtime.tv_sec = VNOVAL;
1114 vap->va_mtime.tv_nsec = VNOVAL;
1115 vap->va_ctime.tv_sec = VNOVAL;
1116 vap->va_ctime.tv_nsec = VNOVAL;
1117 vap->va_birthtime.tv_sec = VNOVAL;
1118 vap->va_birthtime.tv_nsec = VNOVAL;
1119 vap->va_flags = VNOVAL;
1120 vap->va_gen = VNOVAL;
1121 vap->va_vaflags = 0;
1122 }
1123
1124 /*
1125 * Try to reduce the total number of vnodes.
1126 *
1127 * This routine (and its user) are buggy in at least the following ways:
1128 * - all parameters were picked years ago when RAM sizes were significantly
1129 * smaller
1130 * - it can pick vnodes based on pages used by the vm object, but filesystems
1131 * like ZFS don't use it making the pick broken
1132 * - since ZFS has its own aging policy it gets partially combated by this one
1133 * - a dedicated method should be provided for filesystems to let them decide
1134 * whether the vnode should be recycled
1135 *
1136 * This routine is called when we have too many vnodes. It attempts
1137 * to free <count> vnodes and will potentially free vnodes that still
1138 * have VM backing store (VM backing store is typically the cause
1139 * of a vnode blowout so we want to do this). Therefore, this operation
1140 * is not considered cheap.
1141 *
1142 * A number of conditions may prevent a vnode from being reclaimed.
1143 * the buffer cache may have references on the vnode, a directory
1144 * vnode may still have references due to the namei cache representing
1145 * underlying files, or the vnode may be in active use. It is not
1146 * desirable to reuse such vnodes. These conditions may cause the
1147 * number of vnodes to reach some minimum value regardless of what
1148 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low.
1149 *
1150 * @param reclaim_nc_src Only reclaim directories with outgoing namecache
1151 * entries if this argument is strue
1152 * @param trigger Only reclaim vnodes with fewer than this many resident
1153 * pages.
1154 * @param target How many vnodes to reclaim.
1155 * @return The number of vnodes that were reclaimed.
1156 */
1157 static int
1158 vlrureclaim(bool reclaim_nc_src, int trigger, u_long target)
1159 {
1160 struct vnode *vp, *mvp;
1161 struct mount *mp;
1162 struct vm_object *object;
1163 u_long done;
1164 bool retried;
1165
1166 mtx_assert(&vnode_list_mtx, MA_OWNED);
1167
1168 retried = false;
1169 done = 0;
1170
1171 mvp = vnode_list_reclaim_marker;
1172 restart:
1173 vp = mvp;
1174 while (done < target) {
1175 vp = TAILQ_NEXT(vp, v_vnodelist);
1176 if (__predict_false(vp == NULL))
1177 break;
1178
1179 if (__predict_false(vp->v_type == VMARKER))
1180 continue;
1181
1182 /*
1183 * If it's been deconstructed already, it's still
1184 * referenced, or it exceeds the trigger, skip it.
1185 * Also skip free vnodes. We are trying to make space
1186 * to expand the free list, not reduce it.
1187 */
1188 if (vp->v_usecount > 0 || vp->v_holdcnt == 0 ||
1189 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)))
1190 goto next_iter;
1191
1192 if (vp->v_type == VBAD || vp->v_type == VNON)
1193 goto next_iter;
1194
1195 object = atomic_load_ptr(&vp->v_object);
1196 if (object == NULL || object->resident_page_count > trigger) {
1197 goto next_iter;
1198 }
1199
1200 /*
1201 * Handle races against vnode allocation. Filesystems lock the
1202 * vnode some time after it gets returned from getnewvnode,
1203 * despite type and hold count being manipulated earlier.
1204 * Resorting to checking v_mount restores guarantees present
1205 * before the global list was reworked to contain all vnodes.
1206 */
1207 if (!VI_TRYLOCK(vp))
1208 goto next_iter;
1209 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1210 VI_UNLOCK(vp);
1211 goto next_iter;
1212 }
1213 if (vp->v_mount == NULL) {
1214 VI_UNLOCK(vp);
1215 goto next_iter;
1216 }
1217 vholdl(vp);
1218 VI_UNLOCK(vp);
1219 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1220 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1221 mtx_unlock(&vnode_list_mtx);
1222
1223 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
1224 vdrop_recycle(vp);
1225 goto next_iter_unlocked;
1226 }
1227 if (VOP_LOCK(vp, LK_EXCLUSIVE|LK_NOWAIT) != 0) {
1228 vdrop_recycle(vp);
1229 vn_finished_write(mp);
1230 goto next_iter_unlocked;
1231 }
1232
1233 VI_LOCK(vp);
1234 if (vp->v_usecount > 0 ||
1235 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
1236 (vp->v_object != NULL && vp->v_object->handle == vp &&
1237 vp->v_object->resident_page_count > trigger)) {
1238 VOP_UNLOCK(vp);
1239 vdropl_recycle(vp);
1240 vn_finished_write(mp);
1241 goto next_iter_unlocked;
1242 }
1243 counter_u64_add(recycles_count, 1);
1244 vgonel(vp);
1245 VOP_UNLOCK(vp);
1246 vdropl_recycle(vp);
1247 vn_finished_write(mp);
1248 done++;
1249 next_iter_unlocked:
1250 if (should_yield())
1251 kern_yield(PRI_USER);
1252 mtx_lock(&vnode_list_mtx);
1253 goto restart;
1254 next_iter:
1255 MPASS(vp->v_type != VMARKER);
1256 if (!should_yield())
1257 continue;
1258 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1259 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1260 mtx_unlock(&vnode_list_mtx);
1261 kern_yield(PRI_USER);
1262 mtx_lock(&vnode_list_mtx);
1263 goto restart;
1264 }
1265 if (done == 0 && !retried) {
1266 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1267 TAILQ_INSERT_HEAD(&vnode_list, mvp, v_vnodelist);
1268 retried = true;
1269 goto restart;
1270 }
1271 return (done);
1272 }
1273
1274 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
1275 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
1276 0,
1277 "limit on vnode free requests per call to the vnlru_free routine");
1278
1279 /*
1280 * Attempt to reduce the free list by the requested amount.
1281 */
1282 static int
1283 vnlru_free_impl(int count, struct vfsops *mnt_op, struct vnode *mvp)
1284 {
1285 struct vnode *vp;
1286 struct mount *mp;
1287 int ocount;
1288
1289 mtx_assert(&vnode_list_mtx, MA_OWNED);
1290 if (count > max_vnlru_free)
1291 count = max_vnlru_free;
1292 ocount = count;
1293 vp = mvp;
1294 for (;;) {
1295 if (count == 0) {
1296 break;
1297 }
1298 vp = TAILQ_NEXT(vp, v_vnodelist);
1299 if (__predict_false(vp == NULL)) {
1300 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1301 TAILQ_INSERT_TAIL(&vnode_list, mvp, v_vnodelist);
1302 break;
1303 }
1304 if (__predict_false(vp->v_type == VMARKER))
1305 continue;
1306 if (vp->v_holdcnt > 0)
1307 continue;
1308 /*
1309 * Don't recycle if our vnode is from different type
1310 * of mount point. Note that mp is type-safe, the
1311 * check does not reach unmapped address even if
1312 * vnode is reclaimed.
1313 */
1314 if (mnt_op != NULL && (mp = vp->v_mount) != NULL &&
1315 mp->mnt_op != mnt_op) {
1316 continue;
1317 }
1318 if (__predict_false(vp->v_type == VBAD || vp->v_type == VNON)) {
1319 continue;
1320 }
1321 if (!vhold_recycle_free(vp))
1322 continue;
1323 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1324 TAILQ_INSERT_AFTER(&vnode_list, vp, mvp, v_vnodelist);
1325 mtx_unlock(&vnode_list_mtx);
1326 /*
1327 * FIXME: ignores the return value, meaning it may be nothing
1328 * got recycled but it claims otherwise to the caller.
1329 *
1330 * Originally the value started being ignored in 2005 with
1331 * 114a1006a8204aa156e1f9ad6476cdff89cada7f .
1332 *
1333 * Respecting the value can run into significant stalls if most
1334 * vnodes belong to one file system and it has writes
1335 * suspended. In presence of many threads and millions of
1336 * vnodes they keep contending on the vnode_list_mtx lock only
1337 * to find vnodes they can't recycle.
1338 *
1339 * The solution would be to pre-check if the vnode is likely to
1340 * be recycle-able, but it needs to happen with the
1341 * vnode_list_mtx lock held. This runs into a problem where
1342 * VOP_GETWRITEMOUNT (currently needed to find out about if
1343 * writes are frozen) can take locks which LOR against it.
1344 *
1345 * Check nullfs for one example (null_getwritemount).
1346 */
1347 vtryrecycle(vp);
1348 count--;
1349 mtx_lock(&vnode_list_mtx);
1350 vp = mvp;
1351 }
1352 return (ocount - count);
1353 }
1354
1355 static int
1356 vnlru_free_locked(int count)
1357 {
1358
1359 mtx_assert(&vnode_list_mtx, MA_OWNED);
1360 return (vnlru_free_impl(count, NULL, vnode_list_free_marker));
1361 }
1362
1363 void
1364 vnlru_free_vfsops(int count, struct vfsops *mnt_op, struct vnode *mvp)
1365 {
1366
1367 MPASS(mnt_op != NULL);
1368 MPASS(mvp != NULL);
1369 VNPASS(mvp->v_type == VMARKER, mvp);
1370 mtx_lock(&vnode_list_mtx);
1371 vnlru_free_impl(count, mnt_op, mvp);
1372 mtx_unlock(&vnode_list_mtx);
1373 }
1374
1375 struct vnode *
1376 vnlru_alloc_marker(void)
1377 {
1378 struct vnode *mvp;
1379
1380 mvp = vn_alloc_marker(NULL);
1381 mtx_lock(&vnode_list_mtx);
1382 TAILQ_INSERT_BEFORE(vnode_list_free_marker, mvp, v_vnodelist);
1383 mtx_unlock(&vnode_list_mtx);
1384 return (mvp);
1385 }
1386
1387 void
1388 vnlru_free_marker(struct vnode *mvp)
1389 {
1390 mtx_lock(&vnode_list_mtx);
1391 TAILQ_REMOVE(&vnode_list, mvp, v_vnodelist);
1392 mtx_unlock(&vnode_list_mtx);
1393 vn_free_marker(mvp);
1394 }
1395
1396 static void
1397 vnlru_recalc(void)
1398 {
1399
1400 mtx_assert(&vnode_list_mtx, MA_OWNED);
1401 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1402 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1403 vlowat = vhiwat / 2;
1404 }
1405
1406 /*
1407 * Attempt to recycle vnodes in a context that is always safe to block.
1408 * Calling vlrurecycle() from the bowels of filesystem code has some
1409 * interesting deadlock problems.
1410 */
1411 static struct proc *vnlruproc;
1412 static int vnlruproc_sig;
1413
1414 /*
1415 * The main freevnodes counter is only updated when threads requeue their vnode
1416 * batches. CPUs are conditionally walked to compute a more accurate total.
1417 *
1418 * Limit how much of a slop are we willing to tolerate. Note: the actual value
1419 * at any given moment can still exceed slop, but it should not be by significant
1420 * margin in practice.
1421 */
1422 #define VNLRU_FREEVNODES_SLOP 128
1423
1424 static __inline void
1425 vfs_freevnodes_inc(void)
1426 {
1427 struct vdbatch *vd;
1428
1429 critical_enter();
1430 vd = DPCPU_PTR(vd);
1431 vd->freevnodes++;
1432 critical_exit();
1433 }
1434
1435 static __inline void
1436 vfs_freevnodes_dec(void)
1437 {
1438 struct vdbatch *vd;
1439
1440 critical_enter();
1441 vd = DPCPU_PTR(vd);
1442 vd->freevnodes--;
1443 critical_exit();
1444 }
1445
1446 static u_long
1447 vnlru_read_freevnodes(void)
1448 {
1449 struct vdbatch *vd;
1450 long slop;
1451 int cpu;
1452
1453 mtx_assert(&vnode_list_mtx, MA_OWNED);
1454 if (freevnodes > freevnodes_old)
1455 slop = freevnodes - freevnodes_old;
1456 else
1457 slop = freevnodes_old - freevnodes;
1458 if (slop < VNLRU_FREEVNODES_SLOP)
1459 return (freevnodes >= 0 ? freevnodes : 0);
1460 freevnodes_old = freevnodes;
1461 CPU_FOREACH(cpu) {
1462 vd = DPCPU_ID_PTR((cpu), vd);
1463 freevnodes_old += vd->freevnodes;
1464 }
1465 return (freevnodes_old >= 0 ? freevnodes_old : 0);
1466 }
1467
1468 static bool
1469 vnlru_under(u_long rnumvnodes, u_long limit)
1470 {
1471 u_long rfreevnodes, space;
1472
1473 if (__predict_false(rnumvnodes > desiredvnodes))
1474 return (true);
1475
1476 space = desiredvnodes - rnumvnodes;
1477 if (space < limit) {
1478 rfreevnodes = vnlru_read_freevnodes();
1479 if (rfreevnodes > wantfreevnodes)
1480 space += rfreevnodes - wantfreevnodes;
1481 }
1482 return (space < limit);
1483 }
1484
1485 static bool
1486 vnlru_under_unlocked(u_long rnumvnodes, u_long limit)
1487 {
1488 long rfreevnodes, space;
1489
1490 if (__predict_false(rnumvnodes > desiredvnodes))
1491 return (true);
1492
1493 space = desiredvnodes - rnumvnodes;
1494 if (space < limit) {
1495 rfreevnodes = atomic_load_long(&freevnodes);
1496 if (rfreevnodes > wantfreevnodes)
1497 space += rfreevnodes - wantfreevnodes;
1498 }
1499 return (space < limit);
1500 }
1501
1502 static void
1503 vnlru_kick(void)
1504 {
1505
1506 mtx_assert(&vnode_list_mtx, MA_OWNED);
1507 if (vnlruproc_sig == 0) {
1508 vnlruproc_sig = 1;
1509 wakeup(vnlruproc);
1510 }
1511 }
1512
1513 static void
1514 vnlru_proc(void)
1515 {
1516 u_long rnumvnodes, rfreevnodes, target;
1517 unsigned long onumvnodes;
1518 int done, force, trigger, usevnodes;
1519 bool reclaim_nc_src, want_reread;
1520
1521 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1522 SHUTDOWN_PRI_FIRST);
1523
1524 force = 0;
1525 want_reread = false;
1526 for (;;) {
1527 kproc_suspend_check(vnlruproc);
1528 mtx_lock(&vnode_list_mtx);
1529 rnumvnodes = atomic_load_long(&numvnodes);
1530
1531 if (want_reread) {
1532 force = vnlru_under(numvnodes, vhiwat) ? 1 : 0;
1533 want_reread = false;
1534 }
1535
1536 /*
1537 * If numvnodes is too large (due to desiredvnodes being
1538 * adjusted using its sysctl, or emergency growth), first
1539 * try to reduce it by discarding from the free list.
1540 */
1541 if (rnumvnodes > desiredvnodes) {
1542 vnlru_free_locked(rnumvnodes - desiredvnodes);
1543 rnumvnodes = atomic_load_long(&numvnodes);
1544 }
1545 /*
1546 * Sleep if the vnode cache is in a good state. This is
1547 * when it is not over-full and has space for about a 4%
1548 * or 9% expansion (by growing its size or inexcessively
1549 * reducing its free list). Otherwise, try to reclaim
1550 * space for a 10% expansion.
1551 */
1552 if (vstir && force == 0) {
1553 force = 1;
1554 vstir = 0;
1555 }
1556 if (force == 0 && !vnlru_under(rnumvnodes, vlowat)) {
1557 vnlruproc_sig = 0;
1558 wakeup(&vnlruproc_sig);
1559 msleep(vnlruproc, &vnode_list_mtx,
1560 PVFS|PDROP, "vlruwt", hz);
1561 continue;
1562 }
1563 rfreevnodes = vnlru_read_freevnodes();
1564
1565 onumvnodes = rnumvnodes;
1566 /*
1567 * Calculate parameters for recycling. These are the same
1568 * throughout the loop to give some semblance of fairness.
1569 * The trigger point is to avoid recycling vnodes with lots
1570 * of resident pages. We aren't trying to free memory; we
1571 * are trying to recycle or at least free vnodes.
1572 */
1573 if (rnumvnodes <= desiredvnodes)
1574 usevnodes = rnumvnodes - rfreevnodes;
1575 else
1576 usevnodes = rnumvnodes;
1577 if (usevnodes <= 0)
1578 usevnodes = 1;
1579 /*
1580 * The trigger value is chosen to give a conservatively
1581 * large value to ensure that it alone doesn't prevent
1582 * making progress. The value can easily be so large that
1583 * it is effectively infinite in some congested and
1584 * misconfigured cases, and this is necessary. Normally
1585 * it is about 8 to 100 (pages), which is quite large.
1586 */
1587 trigger = vm_cnt.v_page_count * 2 / usevnodes;
1588 if (force < 2)
1589 trigger = vsmalltrigger;
1590 reclaim_nc_src = force >= 3;
1591 target = rnumvnodes * (int64_t)gapvnodes / imax(desiredvnodes, 1);
1592 target = target / 10 + 1;
1593 done = vlrureclaim(reclaim_nc_src, trigger, target);
1594 mtx_unlock(&vnode_list_mtx);
1595 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1596 uma_reclaim(UMA_RECLAIM_DRAIN);
1597 if (done == 0) {
1598 if (force == 0 || force == 1) {
1599 force = 2;
1600 continue;
1601 }
1602 if (force == 2) {
1603 force = 3;
1604 continue;
1605 }
1606 want_reread = true;
1607 force = 0;
1608 vnlru_nowhere++;
1609 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1610 } else {
1611 want_reread = true;
1612 kern_yield(PRI_USER);
1613 }
1614 }
1615 }
1616
1617 static struct kproc_desc vnlru_kp = {
1618 "vnlru",
1619 vnlru_proc,
1620 &vnlruproc
1621 };
1622 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1623 &vnlru_kp);
1624
1625 /*
1626 * Routines having to do with the management of the vnode table.
1627 */
1628
1629 /*
1630 * Try to recycle a freed vnode. We abort if anyone picks up a reference
1631 * before we actually vgone(). This function must be called with the vnode
1632 * held to prevent the vnode from being returned to the free list midway
1633 * through vgone().
1634 */
1635 static int
1636 vtryrecycle(struct vnode *vp)
1637 {
1638 struct mount *vnmp;
1639
1640 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1641 VNASSERT(vp->v_holdcnt, vp,
1642 ("vtryrecycle: Recycling vp %p without a reference.", vp));
1643 /*
1644 * This vnode may found and locked via some other list, if so we
1645 * can't recycle it yet.
1646 */
1647 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1648 CTR2(KTR_VFS,
1649 "%s: impossible to recycle, vp %p lock is already held",
1650 __func__, vp);
1651 vdrop_recycle(vp);
1652 return (EWOULDBLOCK);
1653 }
1654 /*
1655 * Don't recycle if its filesystem is being suspended.
1656 */
1657 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1658 VOP_UNLOCK(vp);
1659 CTR2(KTR_VFS,
1660 "%s: impossible to recycle, cannot start the write for %p",
1661 __func__, vp);
1662 vdrop_recycle(vp);
1663 return (EBUSY);
1664 }
1665 /*
1666 * If we got this far, we need to acquire the interlock and see if
1667 * anyone picked up this vnode from another list. If not, we will
1668 * mark it with DOOMED via vgonel() so that anyone who does find it
1669 * will skip over it.
1670 */
1671 VI_LOCK(vp);
1672 if (vp->v_usecount) {
1673 VOP_UNLOCK(vp);
1674 vdropl_recycle(vp);
1675 vn_finished_write(vnmp);
1676 CTR2(KTR_VFS,
1677 "%s: impossible to recycle, %p is already referenced",
1678 __func__, vp);
1679 return (EBUSY);
1680 }
1681 if (!VN_IS_DOOMED(vp)) {
1682 counter_u64_add(recycles_free_count, 1);
1683 vgonel(vp);
1684 }
1685 VOP_UNLOCK(vp);
1686 vdropl_recycle(vp);
1687 vn_finished_write(vnmp);
1688 return (0);
1689 }
1690
1691 /*
1692 * Allocate a new vnode.
1693 *
1694 * The operation never returns an error. Returning an error was disabled
1695 * in r145385 (dated 2005) with the following comment:
1696 *
1697 * XXX Not all VFS_VGET/ffs_vget callers check returns.
1698 *
1699 * Given the age of this commit (almost 15 years at the time of writing this
1700 * comment) restoring the ability to fail requires a significant audit of
1701 * all codepaths.
1702 *
1703 * The routine can try to free a vnode or stall for up to 1 second waiting for
1704 * vnlru to clear things up, but ultimately always performs a M_WAITOK allocation.
1705 */
1706 static u_long vn_alloc_cyclecount;
1707
1708 static struct vnode * __noinline
1709 vn_alloc_hard(struct mount *mp)
1710 {
1711 u_long rnumvnodes, rfreevnodes;
1712
1713 mtx_lock(&vnode_list_mtx);
1714 rnumvnodes = atomic_load_long(&numvnodes);
1715 if (rnumvnodes + 1 < desiredvnodes) {
1716 vn_alloc_cyclecount = 0;
1717 goto alloc;
1718 }
1719 rfreevnodes = vnlru_read_freevnodes();
1720 if (vn_alloc_cyclecount++ >= rfreevnodes) {
1721 vn_alloc_cyclecount = 0;
1722 vstir = 1;
1723 }
1724 /*
1725 * Grow the vnode cache if it will not be above its target max
1726 * after growing. Otherwise, if the free list is nonempty, try
1727 * to reclaim 1 item from it before growing the cache (possibly
1728 * above its target max if the reclamation failed or is delayed).
1729 * Otherwise, wait for some space. In all cases, schedule
1730 * vnlru_proc() if we are getting short of space. The watermarks
1731 * should be chosen so that we never wait or even reclaim from
1732 * the free list to below its target minimum.
1733 */
1734 if (vnlru_free_locked(1) > 0)
1735 goto alloc;
1736 if (mp == NULL || (mp->mnt_kern_flag & MNTK_SUSPEND) == 0) {
1737 /*
1738 * Wait for space for a new vnode.
1739 */
1740 vnlru_kick();
1741 msleep(&vnlruproc_sig, &vnode_list_mtx, PVFS, "vlruwk", hz);
1742 if (atomic_load_long(&numvnodes) + 1 > desiredvnodes &&
1743 vnlru_read_freevnodes() > 1)
1744 vnlru_free_locked(1);
1745 }
1746 alloc:
1747 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1748 if (vnlru_under(rnumvnodes, vlowat))
1749 vnlru_kick();
1750 mtx_unlock(&vnode_list_mtx);
1751 return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1752 }
1753
1754 static struct vnode *
1755 vn_alloc(struct mount *mp)
1756 {
1757 u_long rnumvnodes;
1758
1759 if (__predict_false(vn_alloc_cyclecount != 0))
1760 return (vn_alloc_hard(mp));
1761 rnumvnodes = atomic_fetchadd_long(&numvnodes, 1) + 1;
1762 if (__predict_false(vnlru_under_unlocked(rnumvnodes, vlowat))) {
1763 atomic_subtract_long(&numvnodes, 1);
1764 return (vn_alloc_hard(mp));
1765 }
1766
1767 return (uma_zalloc_smr(vnode_zone, M_WAITOK));
1768 }
1769
1770 static void
1771 vn_free(struct vnode *vp)
1772 {
1773
1774 atomic_subtract_long(&numvnodes, 1);
1775 uma_zfree_smr(vnode_zone, vp);
1776 }
1777
1778 /*
1779 * Return the next vnode from the free list.
1780 */
1781 int
1782 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1783 struct vnode **vpp)
1784 {
1785 struct vnode *vp;
1786 struct thread *td;
1787 struct lock_object *lo;
1788
1789 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1790
1791 KASSERT(vops->registered,
1792 ("%s: not registered vector op %p\n", __func__, vops));
1793
1794 td = curthread;
1795 if (td->td_vp_reserved != NULL) {
1796 vp = td->td_vp_reserved;
1797 td->td_vp_reserved = NULL;
1798 } else {
1799 vp = vn_alloc(mp);
1800 }
1801 counter_u64_add(vnodes_created, 1);
1802
1803 vn_set_state(vp, VSTATE_UNINITIALIZED);
1804
1805 /*
1806 * Locks are given the generic name "vnode" when created.
1807 * Follow the historic practice of using the filesystem
1808 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1809 *
1810 * Locks live in a witness group keyed on their name. Thus,
1811 * when a lock is renamed, it must also move from the witness
1812 * group of its old name to the witness group of its new name.
1813 *
1814 * The change only needs to be made when the vnode moves
1815 * from one filesystem type to another. We ensure that each
1816 * filesystem use a single static name pointer for its tag so
1817 * that we can compare pointers rather than doing a strcmp().
1818 */
1819 lo = &vp->v_vnlock->lock_object;
1820 #ifdef WITNESS
1821 if (lo->lo_name != tag) {
1822 #endif
1823 lo->lo_name = tag;
1824 #ifdef WITNESS
1825 WITNESS_DESTROY(lo);
1826 WITNESS_INIT(lo, tag);
1827 }
1828 #endif
1829 /*
1830 * By default, don't allow shared locks unless filesystems opt-in.
1831 */
1832 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1833 /*
1834 * Finalize various vnode identity bits.
1835 */
1836 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1837 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1838 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1839 vp->v_type = VNON;
1840 vp->v_op = vops;
1841 vp->v_irflag = 0;
1842 v_init_counters(vp);
1843 vn_seqc_init(vp);
1844 vp->v_bufobj.bo_ops = &buf_ops_bio;
1845 #ifdef DIAGNOSTIC
1846 if (mp == NULL && vops != &dead_vnodeops)
1847 printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1848 #endif
1849 #ifdef MAC
1850 mac_vnode_init(vp);
1851 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1852 mac_vnode_associate_singlelabel(mp, vp);
1853 #endif
1854 if (mp != NULL) {
1855 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1856 }
1857
1858 /*
1859 * For the filesystems which do not use vfs_hash_insert(),
1860 * still initialize v_hash to have vfs_hash_index() useful.
1861 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1862 * its own hashing.
1863 */
1864 vp->v_hash = (uintptr_t)vp >> vnsz2log;
1865
1866 *vpp = vp;
1867 return (0);
1868 }
1869
1870 void
1871 getnewvnode_reserve(void)
1872 {
1873 struct thread *td;
1874
1875 td = curthread;
1876 MPASS(td->td_vp_reserved == NULL);
1877 td->td_vp_reserved = vn_alloc(NULL);
1878 }
1879
1880 void
1881 getnewvnode_drop_reserve(void)
1882 {
1883 struct thread *td;
1884
1885 td = curthread;
1886 if (td->td_vp_reserved != NULL) {
1887 vn_free(td->td_vp_reserved);
1888 td->td_vp_reserved = NULL;
1889 }
1890 }
1891
1892 static void __noinline
1893 freevnode(struct vnode *vp)
1894 {
1895 struct bufobj *bo;
1896
1897 /*
1898 * The vnode has been marked for destruction, so free it.
1899 *
1900 * The vnode will be returned to the zone where it will
1901 * normally remain until it is needed for another vnode. We
1902 * need to cleanup (or verify that the cleanup has already
1903 * been done) any residual data left from its current use
1904 * so as not to contaminate the freshly allocated vnode.
1905 */
1906 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
1907 /*
1908 * Paired with vgone.
1909 */
1910 vn_seqc_write_end_free(vp);
1911
1912 bo = &vp->v_bufobj;
1913 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
1914 VNPASS(vp->v_holdcnt == VHOLD_NO_SMR, vp);
1915 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
1916 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
1917 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
1918 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
1919 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
1920 ("clean blk trie not empty"));
1921 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
1922 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
1923 ("dirty blk trie not empty"));
1924 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
1925 ("Dangling rangelock waiters"));
1926 VNASSERT((vp->v_iflag & (VI_DOINGINACT | VI_OWEINACT)) == 0, vp,
1927 ("Leaked inactivation"));
1928 VI_UNLOCK(vp);
1929 cache_assert_no_entries(vp);
1930
1931 #ifdef MAC
1932 mac_vnode_destroy(vp);
1933 #endif
1934 if (vp->v_pollinfo != NULL) {
1935 /*
1936 * Use LK_NOWAIT to shut up witness about the lock. We may get
1937 * here while having another vnode locked when trying to
1938 * satisfy a lookup and needing to recycle.
1939 */
1940 VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT);
1941 destroy_vpollinfo(vp->v_pollinfo);
1942 VOP_UNLOCK(vp);
1943 vp->v_pollinfo = NULL;
1944 }
1945 vp->v_mountedhere = NULL;
1946 vp->v_unpcb = NULL;
1947 vp->v_rdev = NULL;
1948 vp->v_fifoinfo = NULL;
1949 vp->v_iflag = 0;
1950 vp->v_vflag = 0;
1951 bo->bo_flag = 0;
1952 vn_free(vp);
1953 }
1954
1955 /*
1956 * Delete from old mount point vnode list, if on one.
1957 */
1958 static void
1959 delmntque(struct vnode *vp)
1960 {
1961 struct mount *mp;
1962
1963 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
1964
1965 mp = vp->v_mount;
1966 MNT_ILOCK(mp);
1967 VI_LOCK(vp);
1968 vp->v_mount = NULL;
1969 VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1970 ("bad mount point vnode list size"));
1971 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1972 mp->mnt_nvnodelistsize--;
1973 MNT_REL(mp);
1974 MNT_IUNLOCK(mp);
1975 /*
1976 * The caller expects the interlock to be still held.
1977 */
1978 ASSERT_VI_LOCKED(vp, __func__);
1979 }
1980
1981 static int
1982 insmntque1_int(struct vnode *vp, struct mount *mp, bool dtr)
1983 {
1984
1985 KASSERT(vp->v_mount == NULL,
1986 ("insmntque: vnode already on per mount vnode list"));
1987 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1988 if ((mp->mnt_kern_flag & MNTK_UNLOCKED_INSMNTQUE) == 0) {
1989 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1990 } else {
1991 KASSERT(!dtr,
1992 ("%s: can't have MNTK_UNLOCKED_INSMNTQUE and cleanup",
1993 __func__));
1994 }
1995
1996 /*
1997 * We acquire the vnode interlock early to ensure that the
1998 * vnode cannot be recycled by another process releasing a
1999 * holdcnt on it before we get it on both the vnode list
2000 * and the active vnode list. The mount mutex protects only
2001 * manipulation of the vnode list and the vnode freelist
2002 * mutex protects only manipulation of the active vnode list.
2003 * Hence the need to hold the vnode interlock throughout.
2004 */
2005 MNT_ILOCK(mp);
2006 VI_LOCK(vp);
2007 if (((mp->mnt_kern_flag & MNTK_UNMOUNT) != 0 &&
2008 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
2009 mp->mnt_nvnodelistsize == 0)) &&
2010 (vp->v_vflag & VV_FORCEINSMQ) == 0) {
2011 VI_UNLOCK(vp);
2012 MNT_IUNLOCK(mp);
2013 if (dtr) {
2014 vp->v_data = NULL;
2015 vp->v_op = &dead_vnodeops;
2016 vgone(vp);
2017 vput(vp);
2018 }
2019 return (EBUSY);
2020 }
2021 vp->v_mount = mp;
2022 MNT_REF(mp);
2023 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
2024 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
2025 ("neg mount point vnode list size"));
2026 mp->mnt_nvnodelistsize++;
2027 VI_UNLOCK(vp);
2028 MNT_IUNLOCK(mp);
2029 return (0);
2030 }
2031
2032 /*
2033 * Insert into list of vnodes for the new mount point, if available.
2034 * insmntque() reclaims the vnode on insertion failure, insmntque1()
2035 * leaves handling of the vnode to the caller.
2036 */
2037 int
2038 insmntque(struct vnode *vp, struct mount *mp)
2039 {
2040 return (insmntque1_int(vp, mp, true));
2041 }
2042
2043 int
2044 insmntque1(struct vnode *vp, struct mount *mp)
2045 {
2046 return (insmntque1_int(vp, mp, false));
2047 }
2048
2049 /*
2050 * Flush out and invalidate all buffers associated with a bufobj
2051 * Called with the underlying object locked.
2052 */
2053 int
2054 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
2055 {
2056 int error;
2057
2058 BO_LOCK(bo);
2059 if (flags & V_SAVE) {
2060 error = bufobj_wwait(bo, slpflag, slptimeo);
2061 if (error) {
2062 BO_UNLOCK(bo);
2063 return (error);
2064 }
2065 if (bo->bo_dirty.bv_cnt > 0) {
2066 BO_UNLOCK(bo);
2067 do {
2068 error = BO_SYNC(bo, MNT_WAIT);
2069 } while (error == ERELOOKUP);
2070 if (error != 0)
2071 return (error);
2072 BO_LOCK(bo);
2073 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0) {
2074 BO_UNLOCK(bo);
2075 return (EBUSY);
2076 }
2077 }
2078 }
2079 /*
2080 * If you alter this loop please notice that interlock is dropped and
2081 * reacquired in flushbuflist. Special care is needed to ensure that
2082 * no race conditions occur from this.
2083 */
2084 do {
2085 error = flushbuflist(&bo->bo_clean,
2086 flags, bo, slpflag, slptimeo);
2087 if (error == 0 && !(flags & V_CLEANONLY))
2088 error = flushbuflist(&bo->bo_dirty,
2089 flags, bo, slpflag, slptimeo);
2090 if (error != 0 && error != EAGAIN) {
2091 BO_UNLOCK(bo);
2092 return (error);
2093 }
2094 } while (error != 0);
2095
2096 /*
2097 * Wait for I/O to complete. XXX needs cleaning up. The vnode can
2098 * have write I/O in-progress but if there is a VM object then the
2099 * VM object can also have read-I/O in-progress.
2100 */
2101 do {
2102 bufobj_wwait(bo, 0, 0);
2103 if ((flags & V_VMIO) == 0 && bo->bo_object != NULL) {
2104 BO_UNLOCK(bo);
2105 vm_object_pip_wait_unlocked(bo->bo_object, "bovlbx");
2106 BO_LOCK(bo);
2107 }
2108 } while (bo->bo_numoutput > 0);
2109 BO_UNLOCK(bo);
2110
2111 /*
2112 * Destroy the copy in the VM cache, too.
2113 */
2114 if (bo->bo_object != NULL &&
2115 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
2116 VM_OBJECT_WLOCK(bo->bo_object);
2117 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
2118 OBJPR_CLEANONLY : 0);
2119 VM_OBJECT_WUNLOCK(bo->bo_object);
2120 }
2121
2122 #ifdef INVARIANTS
2123 BO_LOCK(bo);
2124 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
2125 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
2126 bo->bo_clean.bv_cnt > 0))
2127 panic("vinvalbuf: flush failed");
2128 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
2129 bo->bo_dirty.bv_cnt > 0)
2130 panic("vinvalbuf: flush dirty failed");
2131 BO_UNLOCK(bo);
2132 #endif
2133 return (0);
2134 }
2135
2136 /*
2137 * Flush out and invalidate all buffers associated with a vnode.
2138 * Called with the underlying object locked.
2139 */
2140 int
2141 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
2142 {
2143
2144 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2145 ASSERT_VOP_LOCKED(vp, "vinvalbuf");
2146 if (vp->v_object != NULL && vp->v_object->handle != vp)
2147 return (0);
2148 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
2149 }
2150
2151 /*
2152 * Flush out buffers on the specified list.
2153 *
2154 */
2155 static int
2156 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
2157 int slptimeo)
2158 {
2159 struct buf *bp, *nbp;
2160 int retval, error;
2161 daddr_t lblkno;
2162 b_xflags_t xflags;
2163
2164 ASSERT_BO_WLOCKED(bo);
2165
2166 retval = 0;
2167 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
2168 /*
2169 * If we are flushing both V_NORMAL and V_ALT buffers then
2170 * do not skip any buffers. If we are flushing only V_NORMAL
2171 * buffers then skip buffers marked as BX_ALTDATA. If we are
2172 * flushing only V_ALT buffers then skip buffers not marked
2173 * as BX_ALTDATA.
2174 */
2175 if (((flags & (V_NORMAL | V_ALT)) != (V_NORMAL | V_ALT)) &&
2176 (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA) != 0) ||
2177 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0))) {
2178 continue;
2179 }
2180 if (nbp != NULL) {
2181 lblkno = nbp->b_lblkno;
2182 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
2183 }
2184 retval = EAGAIN;
2185 error = BUF_TIMELOCK(bp,
2186 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
2187 "flushbuf", slpflag, slptimeo);
2188 if (error) {
2189 BO_LOCK(bo);
2190 return (error != ENOLCK ? error : EAGAIN);
2191 }
2192 KASSERT(bp->b_bufobj == bo,
2193 ("bp %p wrong b_bufobj %p should be %p",
2194 bp, bp->b_bufobj, bo));
2195 /*
2196 * XXX Since there are no node locks for NFS, I
2197 * believe there is a slight chance that a delayed
2198 * write will occur while sleeping just above, so
2199 * check for it.
2200 */
2201 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
2202 (flags & V_SAVE)) {
2203 bremfree(bp);
2204 bp->b_flags |= B_ASYNC;
2205 bwrite(bp);
2206 BO_LOCK(bo);
2207 return (EAGAIN); /* XXX: why not loop ? */
2208 }
2209 bremfree(bp);
2210 bp->b_flags |= (B_INVAL | B_RELBUF);
2211 bp->b_flags &= ~B_ASYNC;
2212 brelse(bp);
2213 BO_LOCK(bo);
2214 if (nbp == NULL)
2215 break;
2216 nbp = gbincore(bo, lblkno);
2217 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2218 != xflags)
2219 break; /* nbp invalid */
2220 }
2221 return (retval);
2222 }
2223
2224 int
2225 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
2226 {
2227 struct buf *bp;
2228 int error;
2229 daddr_t lblkno;
2230
2231 ASSERT_BO_LOCKED(bo);
2232
2233 for (lblkno = startn;;) {
2234 again:
2235 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
2236 if (bp == NULL || bp->b_lblkno >= endn ||
2237 bp->b_lblkno < startn)
2238 break;
2239 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
2240 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
2241 if (error != 0) {
2242 BO_RLOCK(bo);
2243 if (error == ENOLCK)
2244 goto again;
2245 return (error);
2246 }
2247 KASSERT(bp->b_bufobj == bo,
2248 ("bp %p wrong b_bufobj %p should be %p",
2249 bp, bp->b_bufobj, bo));
2250 lblkno = bp->b_lblkno + 1;
2251 if ((bp->b_flags & B_MANAGED) == 0)
2252 bremfree(bp);
2253 bp->b_flags |= B_RELBUF;
2254 /*
2255 * In the VMIO case, use the B_NOREUSE flag to hint that the
2256 * pages backing each buffer in the range are unlikely to be
2257 * reused. Dirty buffers will have the hint applied once
2258 * they've been written.
2259 */
2260 if ((bp->b_flags & B_VMIO) != 0)
2261 bp->b_flags |= B_NOREUSE;
2262 brelse(bp);
2263 BO_RLOCK(bo);
2264 }
2265 return (0);
2266 }
2267
2268 /*
2269 * Truncate a file's buffer and pages to a specified length. This
2270 * is in lieu of the old vinvalbuf mechanism, which performed unneeded
2271 * sync activity.
2272 */
2273 int
2274 vtruncbuf(struct vnode *vp, off_t length, int blksize)
2275 {
2276 struct buf *bp, *nbp;
2277 struct bufobj *bo;
2278 daddr_t startlbn;
2279
2280 CTR4(KTR_VFS, "%s: vp %p with block %d:%ju", __func__,
2281 vp, blksize, (uintmax_t)length);
2282
2283 /*
2284 * Round up to the *next* lbn.
2285 */
2286 startlbn = howmany(length, blksize);
2287
2288 ASSERT_VOP_LOCKED(vp, "vtruncbuf");
2289
2290 bo = &vp->v_bufobj;
2291 restart_unlocked:
2292 BO_LOCK(bo);
2293
2294 while (v_inval_buf_range_locked(vp, bo, startlbn, INT64_MAX) == EAGAIN)
2295 ;
2296
2297 if (length > 0) {
2298 restartsync:
2299 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2300 if (bp->b_lblkno > 0)
2301 continue;
2302 /*
2303 * Since we hold the vnode lock this should only
2304 * fail if we're racing with the buf daemon.
2305 */
2306 if (BUF_LOCK(bp,
2307 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2308 BO_LOCKPTR(bo)) == ENOLCK)
2309 goto restart_unlocked;
2310
2311 VNASSERT((bp->b_flags & B_DELWRI), vp,
2312 ("buf(%p) on dirty queue without DELWRI", bp));
2313
2314 bremfree(bp);
2315 bawrite(bp);
2316 BO_LOCK(bo);
2317 goto restartsync;
2318 }
2319 }
2320
2321 bufobj_wwait(bo, 0, 0);
2322 BO_UNLOCK(bo);
2323 vnode_pager_setsize(vp, length);
2324
2325 return (0);
2326 }
2327
2328 /*
2329 * Invalidate the cached pages of a file's buffer within the range of block
2330 * numbers [startlbn, endlbn).
2331 */
2332 void
2333 v_inval_buf_range(struct vnode *vp, daddr_t startlbn, daddr_t endlbn,
2334 int blksize)
2335 {
2336 struct bufobj *bo;
2337 off_t start, end;
2338
2339 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range");
2340
2341 start = blksize * startlbn;
2342 end = blksize * endlbn;
2343
2344 bo = &vp->v_bufobj;
2345 BO_LOCK(bo);
2346 MPASS(blksize == bo->bo_bsize);
2347
2348 while (v_inval_buf_range_locked(vp, bo, startlbn, endlbn) == EAGAIN)
2349 ;
2350
2351 BO_UNLOCK(bo);
2352 vn_pages_remove(vp, OFF_TO_IDX(start), OFF_TO_IDX(end + PAGE_SIZE - 1));
2353 }
2354
2355 static int
2356 v_inval_buf_range_locked(struct vnode *vp, struct bufobj *bo,
2357 daddr_t startlbn, daddr_t endlbn)
2358 {
2359 struct buf *bp, *nbp;
2360 bool anyfreed;
2361
2362 ASSERT_VOP_LOCKED(vp, "v_inval_buf_range_locked");
2363 ASSERT_BO_LOCKED(bo);
2364
2365 do {
2366 anyfreed = false;
2367 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
2368 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2369 continue;
2370 if (BUF_LOCK(bp,
2371 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2372 BO_LOCKPTR(bo)) == ENOLCK) {
2373 BO_LOCK(bo);
2374 return (EAGAIN);
2375 }
2376
2377 bremfree(bp);
2378 bp->b_flags |= B_INVAL | B_RELBUF;
2379 bp->b_flags &= ~B_ASYNC;
2380 brelse(bp);
2381 anyfreed = true;
2382
2383 BO_LOCK(bo);
2384 if (nbp != NULL &&
2385 (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
2386 nbp->b_vp != vp ||
2387 (nbp->b_flags & B_DELWRI) != 0))
2388 return (EAGAIN);
2389 }
2390
2391 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
2392 if (bp->b_lblkno < startlbn || bp->b_lblkno >= endlbn)
2393 continue;
2394 if (BUF_LOCK(bp,
2395 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
2396 BO_LOCKPTR(bo)) == ENOLCK) {
2397 BO_LOCK(bo);
2398 return (EAGAIN);
2399 }
2400 bremfree(bp);
2401 bp->b_flags |= B_INVAL | B_RELBUF;
2402 bp->b_flags &= ~B_ASYNC;
2403 brelse(bp);
2404 anyfreed = true;
2405
2406 BO_LOCK(bo);
2407 if (nbp != NULL &&
2408 (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
2409 (nbp->b_vp != vp) ||
2410 (nbp->b_flags & B_DELWRI) == 0))
2411 return (EAGAIN);
2412 }
2413 } while (anyfreed);
2414 return (0);
2415 }
2416
2417 static void
2418 buf_vlist_remove(struct buf *bp)
2419 {
2420 struct bufv *bv;
2421 b_xflags_t flags;
2422
2423 flags = bp->b_xflags;
2424
2425 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
2426 ASSERT_BO_WLOCKED(bp->b_bufobj);
2427 KASSERT((flags & (BX_VNDIRTY | BX_VNCLEAN)) != 0 &&
2428 (flags & (BX_VNDIRTY | BX_VNCLEAN)) != (BX_VNDIRTY | BX_VNCLEAN),
2429 ("%s: buffer %p has invalid queue state", __func__, bp));
2430
2431 if ((flags & BX_VNDIRTY) != 0)
2432 bv = &bp->b_bufobj->bo_dirty;
2433 else
2434 bv = &bp->b_bufobj->bo_clean;
2435 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
2436 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
2437 bv->bv_cnt--;
2438 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
2439 }
2440
2441 /*
2442 * Add the buffer to the sorted clean or dirty block list.
2443 *
2444 * NOTE: xflags is passed as a constant, optimizing this inline function!
2445 */
2446 static void
2447 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
2448 {
2449 struct bufv *bv;
2450 struct buf *n;
2451 int error;
2452
2453 ASSERT_BO_WLOCKED(bo);
2454 KASSERT((bo->bo_flag & BO_NOBUFS) == 0,
2455 ("buf_vlist_add: bo %p does not allow bufs", bo));
2456 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
2457 ("dead bo %p", bo));
2458 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
2459 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
2460 bp->b_xflags |= xflags;
2461 if (xflags & BX_VNDIRTY)
2462 bv = &bo->bo_dirty;
2463 else
2464 bv = &bo->bo_clean;
2465
2466 /*
2467 * Keep the list ordered. Optimize empty list insertion. Assume
2468 * we tend to grow at the tail so lookup_le should usually be cheaper
2469 * than _ge.
2470 */
2471 if (bv->bv_cnt == 0 ||
2472 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
2473 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
2474 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
2475 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
2476 else
2477 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
2478 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
2479 if (error)
2480 panic("buf_vlist_add: Preallocated nodes insufficient.");
2481 bv->bv_cnt++;
2482 }
2483
2484 /*
2485 * Look up a buffer using the buffer tries.
2486 */
2487 struct buf *
2488 gbincore(struct bufobj *bo, daddr_t lblkno)
2489 {
2490 struct buf *bp;
2491
2492 ASSERT_BO_LOCKED(bo);
2493 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
2494 if (bp != NULL)
2495 return (bp);
2496 return (BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno));
2497 }
2498
2499 /*
2500 * Look up a buf using the buffer tries, without the bufobj lock. This relies
2501 * on SMR for safe lookup, and bufs being in a no-free zone to provide type
2502 * stability of the result. Like other lockless lookups, the found buf may
2503 * already be invalid by the time this function returns.
2504 */
2505 struct buf *
2506 gbincore_unlocked(struct bufobj *bo, daddr_t lblkno)
2507 {
2508 struct buf *bp;
2509
2510 ASSERT_BO_UNLOCKED(bo);
2511 bp = BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_clean.bv_root, lblkno);
2512 if (bp != NULL)
2513 return (bp);
2514 return (BUF_PCTRIE_LOOKUP_UNLOCKED(&bo->bo_dirty.bv_root, lblkno));
2515 }
2516
2517 /*
2518 * Associate a buffer with a vnode.
2519 */
2520 void
2521 bgetvp(struct vnode *vp, struct buf *bp)
2522 {
2523 struct bufobj *bo;
2524
2525 bo = &vp->v_bufobj;
2526 ASSERT_BO_WLOCKED(bo);
2527 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
2528
2529 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
2530 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
2531 ("bgetvp: bp already attached! %p", bp));
2532
2533 vhold(vp);
2534 bp->b_vp = vp;
2535 bp->b_bufobj = bo;
2536 /*
2537 * Insert onto list for new vnode.
2538 */
2539 buf_vlist_add(bp, bo, BX_VNCLEAN);
2540 }
2541
2542 /*
2543 * Disassociate a buffer from a vnode.
2544 */
2545 void
2546 brelvp(struct buf *bp)
2547 {
2548 struct bufobj *bo;
2549 struct vnode *vp;
2550
2551 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
2552 KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
2553
2554 /*
2555 * Delete from old vnode list, if on one.
2556 */
2557 vp = bp->b_vp; /* XXX */
2558 bo = bp->b_bufobj;
2559 BO_LOCK(bo);
2560 buf_vlist_remove(bp);
2561 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2562 bo->bo_flag &= ~BO_ONWORKLST;
2563 mtx_lock(&sync_mtx);
2564 LIST_REMOVE(bo, bo_synclist);
2565 syncer_worklist_len--;
2566 mtx_unlock(&sync_mtx);
2567 }
2568 bp->b_vp = NULL;
2569 bp->b_bufobj = NULL;
2570 BO_UNLOCK(bo);
2571 vdrop(vp);
2572 }
2573
2574 /*
2575 * Add an item to the syncer work queue.
2576 */
2577 static void
2578 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2579 {
2580 int slot;
2581
2582 ASSERT_BO_WLOCKED(bo);
2583
2584 mtx_lock(&sync_mtx);
2585 if (bo->bo_flag & BO_ONWORKLST)
2586 LIST_REMOVE(bo, bo_synclist);
2587 else {
2588 bo->bo_flag |= BO_ONWORKLST;
2589 syncer_worklist_len++;
2590 }
2591
2592 if (delay > syncer_maxdelay - 2)
2593 delay = syncer_maxdelay - 2;
2594 slot = (syncer_delayno + delay) & syncer_mask;
2595
2596 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2597 mtx_unlock(&sync_mtx);
2598 }
2599
2600 static int
2601 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2602 {
2603 int error, len;
2604
2605 mtx_lock(&sync_mtx);
2606 len = syncer_worklist_len - sync_vnode_count;
2607 mtx_unlock(&sync_mtx);
2608 error = SYSCTL_OUT(req, &len, sizeof(len));
2609 return (error);
2610 }
2611
2612 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len,
2613 CTLTYPE_INT | CTLFLAG_MPSAFE| CTLFLAG_RD, NULL, 0,
2614 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2615
2616 static struct proc *updateproc;
2617 static void sched_sync(void);
2618 static struct kproc_desc up_kp = {
2619 "syncer",
2620 sched_sync,
2621 &updateproc
2622 };
2623 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2624
2625 static int
2626 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2627 {
2628 struct vnode *vp;
2629 struct mount *mp;
2630
2631 *bo = LIST_FIRST(slp);
2632 if (*bo == NULL)
2633 return (0);
2634 vp = bo2vnode(*bo);
2635 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2636 return (1);
2637 /*
2638 * We use vhold in case the vnode does not
2639 * successfully sync. vhold prevents the vnode from
2640 * going away when we unlock the sync_mtx so that
2641 * we can acquire the vnode interlock.
2642 */
2643 vholdl(vp);
2644 mtx_unlock(&sync_mtx);
2645 VI_UNLOCK(vp);
2646 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2647 vdrop(vp);
2648 mtx_lock(&sync_mtx);
2649 return (*bo == LIST_FIRST(slp));
2650 }
2651 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2652 (void) VOP_FSYNC(vp, MNT_LAZY, td);
2653 VOP_UNLOCK(vp);
2654 vn_finished_write(mp);
2655 BO_LOCK(*bo);
2656 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2657 /*
2658 * Put us back on the worklist. The worklist
2659 * routine will remove us from our current
2660 * position and then add us back in at a later
2661 * position.
2662 */
2663 vn_syncer_add_to_worklist(*bo, syncdelay);
2664 }
2665 BO_UNLOCK(*bo);
2666 vdrop(vp);
2667 mtx_lock(&sync_mtx);
2668 return (0);
2669 }
2670
2671 static int first_printf = 1;
2672
2673 /*
2674 * System filesystem synchronizer daemon.
2675 */
2676 static void
2677 sched_sync(void)
2678 {
2679 struct synclist *next, *slp;
2680 struct bufobj *bo;
2681 long starttime;
2682 struct thread *td = curthread;
2683 int last_work_seen;
2684 int net_worklist_len;
2685 int syncer_final_iter;
2686 int error;
2687
2688 last_work_seen = 0;
2689 syncer_final_iter = 0;
2690 syncer_state = SYNCER_RUNNING;
2691 starttime = time_uptime;
2692 td->td_pflags |= TDP_NORUNNINGBUF;
2693
2694 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2695 SHUTDOWN_PRI_LAST);
2696
2697 mtx_lock(&sync_mtx);
2698 for (;;) {
2699 if (syncer_state == SYNCER_FINAL_DELAY &&
2700 syncer_final_iter == 0) {
2701 mtx_unlock(&sync_mtx);
2702 kproc_suspend_check(td->td_proc);
2703 mtx_lock(&sync_mtx);
2704 }
2705 net_worklist_len = syncer_worklist_len - sync_vnode_count;
2706 if (syncer_state != SYNCER_RUNNING &&
2707 starttime != time_uptime) {
2708 if (first_printf) {
2709 printf("\nSyncing disks, vnodes remaining... ");
2710 first_printf = 0;
2711 }
2712 printf("%d ", net_worklist_len);
2713 }
2714 starttime = time_uptime;
2715
2716 /*
2717 * Push files whose dirty time has expired. Be careful
2718 * of interrupt race on slp queue.
2719 *
2720 * Skip over empty worklist slots when shutting down.
2721 */
2722 do {
2723 slp = &syncer_workitem_pending[syncer_delayno];
2724 syncer_delayno += 1;
2725 if (syncer_delayno == syncer_maxdelay)
2726 syncer_delayno = 0;
2727 next = &syncer_workitem_pending[syncer_delayno];
2728 /*
2729 * If the worklist has wrapped since the
2730 * it was emptied of all but syncer vnodes,
2731 * switch to the FINAL_DELAY state and run
2732 * for one more second.
2733 */
2734 if (syncer_state == SYNCER_SHUTTING_DOWN &&
2735 net_worklist_len == 0 &&
2736 last_work_seen == syncer_delayno) {
2737 syncer_state = SYNCER_FINAL_DELAY;
2738 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2739 }
2740 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2741 syncer_worklist_len > 0);
2742
2743 /*
2744 * Keep track of the last time there was anything
2745 * on the worklist other than syncer vnodes.
2746 * Return to the SHUTTING_DOWN state if any
2747 * new work appears.
2748 */
2749 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2750 last_work_seen = syncer_delayno;
2751 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2752 syncer_state = SYNCER_SHUTTING_DOWN;
2753 while (!LIST_EMPTY(slp)) {
2754 error = sync_vnode(slp, &bo, td);
2755 if (error == 1) {
2756 LIST_REMOVE(bo, bo_synclist);
2757 LIST_INSERT_HEAD(next, bo, bo_synclist);
2758 continue;
2759 }
2760
2761 if (first_printf == 0) {
2762 /*
2763 * Drop the sync mutex, because some watchdog
2764 * drivers need to sleep while patting
2765 */
2766 mtx_unlock(&sync_mtx);
2767 wdog_kern_pat(WD_LASTVAL);
2768 mtx_lock(&sync_mtx);
2769 }
2770 }
2771 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2772 syncer_final_iter--;
2773 /*
2774 * The variable rushjob allows the kernel to speed up the
2775 * processing of the filesystem syncer process. A rushjob
2776 * value of N tells the filesystem syncer to process the next
2777 * N seconds worth of work on its queue ASAP. Currently rushjob
2778 * is used by the soft update code to speed up the filesystem
2779 * syncer process when the incore state is getting so far
2780 * ahead of the disk that the kernel memory pool is being
2781 * threatened with exhaustion.
2782 */
2783 if (rushjob > 0) {
2784 rushjob -= 1;
2785 continue;
2786 }
2787 /*
2788 * Just sleep for a short period of time between
2789 * iterations when shutting down to allow some I/O
2790 * to happen.
2791 *
2792 * If it has taken us less than a second to process the
2793 * current work, then wait. Otherwise start right over
2794 * again. We can still lose time if any single round
2795 * takes more than two seconds, but it does not really
2796 * matter as we are just trying to generally pace the
2797 * filesystem activity.
2798 */
2799 if (syncer_state != SYNCER_RUNNING ||
2800 time_uptime == starttime) {
2801 thread_lock(td);
2802 sched_prio(td, PPAUSE);
2803 thread_unlock(td);
2804 }
2805 if (syncer_state != SYNCER_RUNNING)
2806 cv_timedwait(&sync_wakeup, &sync_mtx,
2807 hz / SYNCER_SHUTDOWN_SPEEDUP);
2808 else if (time_uptime == starttime)
2809 cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2810 }
2811 }
2812
2813 /*
2814 * Request the syncer daemon to speed up its work.
2815 * We never push it to speed up more than half of its
2816 * normal turn time, otherwise it could take over the cpu.
2817 */
2818 int
2819 speedup_syncer(void)
2820 {
2821 int ret = 0;
2822
2823 mtx_lock(&sync_mtx);
2824 if (rushjob < syncdelay / 2) {
2825 rushjob += 1;
2826 stat_rush_requests += 1;
2827 ret = 1;
2828 }
2829 mtx_unlock(&sync_mtx);
2830 cv_broadcast(&sync_wakeup);
2831 return (ret);
2832 }
2833
2834 /*
2835 * Tell the syncer to speed up its work and run though its work
2836 * list several times, then tell it to shut down.
2837 */
2838 static void
2839 syncer_shutdown(void *arg, int howto)
2840 {
2841
2842 if (howto & RB_NOSYNC)
2843 return;
2844 mtx_lock(&sync_mtx);
2845 syncer_state = SYNCER_SHUTTING_DOWN;
2846 rushjob = 0;
2847 mtx_unlock(&sync_mtx);
2848 cv_broadcast(&sync_wakeup);
2849 kproc_shutdown(arg, howto);
2850 }
2851
2852 void
2853 syncer_suspend(void)
2854 {
2855
2856 syncer_shutdown(updateproc, 0);
2857 }
2858
2859 void
2860 syncer_resume(void)
2861 {
2862
2863 mtx_lock(&sync_mtx);
2864 first_printf = 1;
2865 syncer_state = SYNCER_RUNNING;
2866 mtx_unlock(&sync_mtx);
2867 cv_broadcast(&sync_wakeup);
2868 kproc_resume(updateproc);
2869 }
2870
2871 /*
2872 * Move the buffer between the clean and dirty lists of its vnode.
2873 */
2874 void
2875 reassignbuf(struct buf *bp)
2876 {
2877 struct vnode *vp;
2878 struct bufobj *bo;
2879 int delay;
2880 #ifdef INVARIANTS
2881 struct bufv *bv;
2882 #endif
2883
2884 vp = bp->b_vp;
2885 bo = bp->b_bufobj;
2886
2887 KASSERT((bp->b_flags & B_PAGING) == 0,
2888 ("%s: cannot reassign paging buffer %p", __func__, bp));
2889
2890 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2891 bp, bp->b_vp, bp->b_flags);
2892
2893 BO_LOCK(bo);
2894 buf_vlist_remove(bp);
2895
2896 /*
2897 * If dirty, put on list of dirty buffers; otherwise insert onto list
2898 * of clean buffers.
2899 */
2900 if (bp->b_flags & B_DELWRI) {
2901 if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2902 switch (vp->v_type) {
2903 case VDIR:
2904 delay = dirdelay;
2905 break;
2906 case VCHR:
2907 delay = metadelay;
2908 break;
2909 default:
2910 delay = filedelay;
2911 }
2912 vn_syncer_add_to_worklist(bo, delay);
2913 }
2914 buf_vlist_add(bp, bo, BX_VNDIRTY);
2915 } else {
2916 buf_vlist_add(bp, bo, BX_VNCLEAN);
2917
2918 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2919 mtx_lock(&sync_mtx);
2920 LIST_REMOVE(bo, bo_synclist);
2921 syncer_worklist_len--;
2922 mtx_unlock(&sync_mtx);
2923 bo->bo_flag &= ~BO_ONWORKLST;
2924 }
2925 }
2926 #ifdef INVARIANTS
2927 bv = &bo->bo_clean;
2928 bp = TAILQ_FIRST(&bv->bv_hd);
2929 KASSERT(bp == NULL || bp->b_bufobj == bo,
2930 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2931 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2932 KASSERT(bp == NULL || bp->b_bufobj == bo,
2933 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2934 bv = &bo->bo_dirty;
2935 bp = TAILQ_FIRST(&bv->bv_hd);
2936 KASSERT(bp == NULL || bp->b_bufobj == bo,
2937 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2938 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2939 KASSERT(bp == NULL || bp->b_bufobj == bo,
2940 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2941 #endif
2942 BO_UNLOCK(bo);
2943 }
2944
2945 static void
2946 v_init_counters(struct vnode *vp)
2947 {
2948
2949 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2950 vp, ("%s called for an initialized vnode", __FUNCTION__));
2951 ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2952
2953 refcount_init(&vp->v_holdcnt, 1);
2954 refcount_init(&vp->v_usecount, 1);
2955 }
2956
2957 /*
2958 * Grab a particular vnode from the free list, increment its
2959 * reference count and lock it. VIRF_DOOMED is set if the vnode
2960 * is being destroyed. Only callers who specify LK_RETRY will
2961 * see doomed vnodes. If inactive processing was delayed in
2962 * vput try to do it here.
2963 *
2964 * usecount is manipulated using atomics without holding any locks.
2965 *
2966 * holdcnt can be manipulated using atomics without holding any locks,
2967 * except when transitioning 1<->0, in which case the interlock is held.
2968 *
2969 * Consumers which don't guarantee liveness of the vnode can use SMR to
2970 * try to get a reference. Note this operation can fail since the vnode
2971 * may be awaiting getting freed by the time they get to it.
2972 */
2973 enum vgetstate
2974 vget_prep_smr(struct vnode *vp)
2975 {
2976 enum vgetstate vs;
2977
2978 VFS_SMR_ASSERT_ENTERED();
2979
2980 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2981 vs = VGET_USECOUNT;
2982 } else {
2983 if (vhold_smr(vp))
2984 vs = VGET_HOLDCNT;
2985 else
2986 vs = VGET_NONE;
2987 }
2988 return (vs);
2989 }
2990
2991 enum vgetstate
2992 vget_prep(struct vnode *vp)
2993 {
2994 enum vgetstate vs;
2995
2996 if (refcount_acquire_if_not_zero(&vp->v_usecount)) {
2997 vs = VGET_USECOUNT;
2998 } else {
2999 vhold(vp);
3000 vs = VGET_HOLDCNT;
3001 }
3002 return (vs);
3003 }
3004
3005 void
3006 vget_abort(struct vnode *vp, enum vgetstate vs)
3007 {
3008
3009 switch (vs) {
3010 case VGET_USECOUNT:
3011 vrele(vp);
3012 break;
3013 case VGET_HOLDCNT:
3014 vdrop(vp);
3015 break;
3016 default:
3017 __assert_unreachable();
3018 }
3019 }
3020
3021 int
3022 vget(struct vnode *vp, int flags)
3023 {
3024 enum vgetstate vs;
3025
3026 vs = vget_prep(vp);
3027 return (vget_finish(vp, flags, vs));
3028 }
3029
3030 int
3031 vget_finish(struct vnode *vp, int flags, enum vgetstate vs)
3032 {
3033 int error;
3034
3035 if ((flags & LK_INTERLOCK) != 0)
3036 ASSERT_VI_LOCKED(vp, __func__);
3037 else
3038 ASSERT_VI_UNLOCKED(vp, __func__);
3039 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3040 VNPASS(vp->v_holdcnt > 0, vp);
3041 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3042
3043 error = vn_lock(vp, flags);
3044 if (__predict_false(error != 0)) {
3045 vget_abort(vp, vs);
3046 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
3047 vp);
3048 return (error);
3049 }
3050
3051 vget_finish_ref(vp, vs);
3052 return (0);
3053 }
3054
3055 void
3056 vget_finish_ref(struct vnode *vp, enum vgetstate vs)
3057 {
3058 int old;
3059
3060 VNPASS(vs == VGET_HOLDCNT || vs == VGET_USECOUNT, vp);
3061 VNPASS(vp->v_holdcnt > 0, vp);
3062 VNPASS(vs == VGET_HOLDCNT || vp->v_usecount > 0, vp);
3063
3064 if (vs == VGET_USECOUNT)
3065 return;
3066
3067 /*
3068 * We hold the vnode. If the usecount is 0 it will be utilized to keep
3069 * the vnode around. Otherwise someone else lended their hold count and
3070 * we have to drop ours.
3071 */
3072 old = atomic_fetchadd_int(&vp->v_usecount, 1);
3073 VNASSERT(old >= 0, vp, ("%s: wrong use count %d", __func__, old));
3074 if (old != 0) {
3075 #ifdef INVARIANTS
3076 old = atomic_fetchadd_int(&vp->v_holdcnt, -1);
3077 VNASSERT(old > 1, vp, ("%s: wrong hold count %d", __func__, old));
3078 #else
3079 refcount_release(&vp->v_holdcnt);
3080 #endif
3081 }
3082 }
3083
3084 void
3085 vref(struct vnode *vp)
3086 {
3087 enum vgetstate vs;
3088
3089 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3090 vs = vget_prep(vp);
3091 vget_finish_ref(vp, vs);
3092 }
3093
3094 void
3095 vrefact(struct vnode *vp)
3096 {
3097
3098 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3099 #ifdef INVARIANTS
3100 int old = atomic_fetchadd_int(&vp->v_usecount, 1);
3101 VNASSERT(old > 0, vp, ("%s: wrong use count %d", __func__, old));
3102 #else
3103 refcount_acquire(&vp->v_usecount);
3104 #endif
3105 }
3106
3107 void
3108 vlazy(struct vnode *vp)
3109 {
3110 struct mount *mp;
3111
3112 VNASSERT(vp->v_holdcnt > 0, vp, ("%s: vnode not held", __func__));
3113
3114 if ((vp->v_mflag & VMP_LAZYLIST) != 0)
3115 return;
3116 /*
3117 * We may get here for inactive routines after the vnode got doomed.
3118 */
3119 if (VN_IS_DOOMED(vp))
3120 return;
3121 mp = vp->v_mount;
3122 mtx_lock(&mp->mnt_listmtx);
3123 if ((vp->v_mflag & VMP_LAZYLIST) == 0) {
3124 vp->v_mflag |= VMP_LAZYLIST;
3125 TAILQ_INSERT_TAIL(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3126 mp->mnt_lazyvnodelistsize++;
3127 }
3128 mtx_unlock(&mp->mnt_listmtx);
3129 }
3130
3131 static void
3132 vunlazy(struct vnode *vp)
3133 {
3134 struct mount *mp;
3135
3136 ASSERT_VI_LOCKED(vp, __func__);
3137 VNPASS(!VN_IS_DOOMED(vp), vp);
3138
3139 mp = vp->v_mount;
3140 mtx_lock(&mp->mnt_listmtx);
3141 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3142 /*
3143 * Don't remove the vnode from the lazy list if another thread
3144 * has increased the hold count. It may have re-enqueued the
3145 * vnode to the lazy list and is now responsible for its
3146 * removal.
3147 */
3148 if (vp->v_holdcnt == 0) {
3149 vp->v_mflag &= ~VMP_LAZYLIST;
3150 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3151 mp->mnt_lazyvnodelistsize--;
3152 }
3153 mtx_unlock(&mp->mnt_listmtx);
3154 }
3155
3156 /*
3157 * This routine is only meant to be called from vgonel prior to dooming
3158 * the vnode.
3159 */
3160 static void
3161 vunlazy_gone(struct vnode *vp)
3162 {
3163 struct mount *mp;
3164
3165 ASSERT_VOP_ELOCKED(vp, __func__);
3166 ASSERT_VI_LOCKED(vp, __func__);
3167 VNPASS(!VN_IS_DOOMED(vp), vp);
3168
3169 if (vp->v_mflag & VMP_LAZYLIST) {
3170 mp = vp->v_mount;
3171 mtx_lock(&mp->mnt_listmtx);
3172 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
3173 vp->v_mflag &= ~VMP_LAZYLIST;
3174 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, vp, v_lazylist);
3175 mp->mnt_lazyvnodelistsize--;
3176 mtx_unlock(&mp->mnt_listmtx);
3177 }
3178 }
3179
3180 static void
3181 vdefer_inactive(struct vnode *vp)
3182 {
3183
3184 ASSERT_VI_LOCKED(vp, __func__);
3185 VNASSERT(vp->v_holdcnt > 0, vp,
3186 ("%s: vnode without hold count", __func__));
3187 if (VN_IS_DOOMED(vp)) {
3188 vdropl(vp);
3189 return;
3190 }
3191 if (vp->v_iflag & VI_DEFINACT) {
3192 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
3193 vdropl(vp);
3194 return;
3195 }
3196 if (vp->v_usecount > 0) {
3197 vp->v_iflag &= ~VI_OWEINACT;
3198 vdropl(vp);
3199 return;
3200 }
3201 vlazy(vp);
3202 vp->v_iflag |= VI_DEFINACT;
3203 VI_UNLOCK(vp);
3204 counter_u64_add(deferred_inact, 1);
3205 }
3206
3207 static void
3208 vdefer_inactive_unlocked(struct vnode *vp)
3209 {
3210
3211 VI_LOCK(vp);
3212 if ((vp->v_iflag & VI_OWEINACT) == 0) {
3213 vdropl(vp);
3214 return;
3215 }
3216 vdefer_inactive(vp);
3217 }
3218
3219 enum vput_op { VRELE, VPUT, VUNREF };
3220
3221 /*
3222 * Handle ->v_usecount transitioning to 0.
3223 *
3224 * By releasing the last usecount we take ownership of the hold count which
3225 * provides liveness of the vnode, meaning we have to vdrop.
3226 *
3227 * For all vnodes we may need to perform inactive processing. It requires an
3228 * exclusive lock on the vnode, while it is legal to call here with only a
3229 * shared lock (or no locks). If locking the vnode in an expected manner fails,
3230 * inactive processing gets deferred to the syncer.
3231 *
3232 * XXX Some filesystems pass in an exclusively locked vnode and strongly depend
3233 * on the lock being held all the way until VOP_INACTIVE. This in particular
3234 * happens with UFS which adds half-constructed vnodes to the hash, where they
3235 * can be found by other code.
3236 */
3237 static void
3238 vput_final(struct vnode *vp, enum vput_op func)
3239 {
3240 int error;
3241 bool want_unlock;
3242
3243 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3244 VNPASS(vp->v_holdcnt > 0, vp);
3245
3246 VI_LOCK(vp);
3247
3248 /*
3249 * By the time we got here someone else might have transitioned
3250 * the count back to > 0.
3251 */
3252 if (vp->v_usecount > 0)
3253 goto out;
3254
3255 /*
3256 * If the vnode is doomed vgone already performed inactive processing
3257 * (if needed).
3258 */
3259 if (VN_IS_DOOMED(vp))
3260 goto out;
3261
3262 if (__predict_true(VOP_NEED_INACTIVE(vp) == 0))
3263 goto out;
3264
3265 if (vp->v_iflag & VI_DOINGINACT)
3266 goto out;
3267
3268 /*
3269 * Locking operations here will drop the interlock and possibly the
3270 * vnode lock, opening a window where the vnode can get doomed all the
3271 * while ->v_usecount is 0. Set VI_OWEINACT to let vgone know to
3272 * perform inactive.
3273 */
3274 vp->v_iflag |= VI_OWEINACT;
3275 want_unlock = false;
3276 error = 0;
3277 switch (func) {
3278 case VRELE:
3279 switch (VOP_ISLOCKED(vp)) {
3280 case LK_EXCLUSIVE:
3281 break;
3282 case LK_EXCLOTHER:
3283 case 0:
3284 want_unlock = true;
3285 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
3286 VI_LOCK(vp);
3287 break;
3288 default:
3289 /*
3290 * The lock has at least one sharer, but we have no way
3291 * to conclude whether this is us. Play it safe and
3292 * defer processing.
3293 */
3294 error = EAGAIN;
3295 break;
3296 }
3297 break;
3298 case VPUT:
3299 want_unlock = true;
3300 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3301 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
3302 LK_NOWAIT);
3303 VI_LOCK(vp);
3304 }
3305 break;
3306 case VUNREF:
3307 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
3308 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
3309 VI_LOCK(vp);
3310 }
3311 break;
3312 }
3313 if (error == 0) {
3314 if (func == VUNREF) {
3315 VNASSERT((vp->v_vflag & VV_UNREF) == 0, vp,
3316 ("recursive vunref"));
3317 vp->v_vflag |= VV_UNREF;
3318 }
3319 for (;;) {
3320 error = vinactive(vp);
3321 if (want_unlock)
3322 VOP_UNLOCK(vp);
3323 if (error != ERELOOKUP || !want_unlock)
3324 break;
3325 VOP_LOCK(vp, LK_EXCLUSIVE);
3326 }
3327 if (func == VUNREF)
3328 vp->v_vflag &= ~VV_UNREF;
3329 vdropl(vp);
3330 } else {
3331 vdefer_inactive(vp);
3332 }
3333 return;
3334 out:
3335 if (func == VPUT)
3336 VOP_UNLOCK(vp);
3337 vdropl(vp);
3338 }
3339
3340 /*
3341 * Decrement ->v_usecount for a vnode.
3342 *
3343 * Releasing the last use count requires additional processing, see vput_final
3344 * above for details.
3345 *
3346 * Comment above each variant denotes lock state on entry and exit.
3347 */
3348
3349 /*
3350 * in: any
3351 * out: same as passed in
3352 */
3353 void
3354 vrele(struct vnode *vp)
3355 {
3356
3357 ASSERT_VI_UNLOCKED(vp, __func__);
3358 if (!refcount_release(&vp->v_usecount))
3359 return;
3360 vput_final(vp, VRELE);
3361 }
3362
3363 /*
3364 * in: locked
3365 * out: unlocked
3366 */
3367 void
3368 vput(struct vnode *vp)
3369 {
3370
3371 ASSERT_VOP_LOCKED(vp, __func__);
3372 ASSERT_VI_UNLOCKED(vp, __func__);
3373 if (!refcount_release(&vp->v_usecount)) {
3374 VOP_UNLOCK(vp);
3375 return;
3376 }
3377 vput_final(vp, VPUT);
3378 }
3379
3380 /*
3381 * in: locked
3382 * out: locked
3383 */
3384 void
3385 vunref(struct vnode *vp)
3386 {
3387
3388 ASSERT_VOP_LOCKED(vp, __func__);
3389 ASSERT_VI_UNLOCKED(vp, __func__);
3390 if (!refcount_release(&vp->v_usecount))
3391 return;
3392 vput_final(vp, VUNREF);
3393 }
3394
3395 void
3396 vhold(struct vnode *vp)
3397 {
3398 int old;
3399
3400 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3401 old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3402 VNASSERT(old >= 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3403 ("%s: wrong hold count %d", __func__, old));
3404 if (old == 0)
3405 vfs_freevnodes_dec();
3406 }
3407
3408 void
3409 vholdnz(struct vnode *vp)
3410 {
3411
3412 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3413 #ifdef INVARIANTS
3414 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3415 VNASSERT(old > 0 && (old & VHOLD_ALL_FLAGS) == 0, vp,
3416 ("%s: wrong hold count %d", __func__, old));
3417 #else
3418 atomic_add_int(&vp->v_holdcnt, 1);
3419 #endif
3420 }
3421
3422 /*
3423 * Grab a hold count unless the vnode is freed.
3424 *
3425 * Only use this routine if vfs smr is the only protection you have against
3426 * freeing the vnode.
3427 *
3428 * The code loops trying to add a hold count as long as the VHOLD_NO_SMR flag
3429 * is not set. After the flag is set the vnode becomes immutable to anyone but
3430 * the thread which managed to set the flag.
3431 *
3432 * It may be tempting to replace the loop with:
3433 * count = atomic_fetchadd_int(&vp->v_holdcnt, 1);
3434 * if (count & VHOLD_NO_SMR) {
3435 * backpedal and error out;
3436 * }
3437 *
3438 * However, while this is more performant, it hinders debugging by eliminating
3439 * the previously mentioned invariant.
3440 */
3441 bool
3442 vhold_smr(struct vnode *vp)
3443 {
3444 int count;
3445
3446 VFS_SMR_ASSERT_ENTERED();
3447
3448 count = atomic_load_int(&vp->v_holdcnt);
3449 for (;;) {
3450 if (count & VHOLD_NO_SMR) {
3451 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3452 ("non-zero hold count with flags %d\n", count));
3453 return (false);
3454 }
3455 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3456 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3457 if (count == 0)
3458 vfs_freevnodes_dec();
3459 return (true);
3460 }
3461 }
3462 }
3463
3464 /*
3465 * Hold a free vnode for recycling.
3466 *
3467 * Note: vnode_init references this comment.
3468 *
3469 * Attempts to recycle only need the global vnode list lock and have no use for
3470 * SMR.
3471 *
3472 * However, vnodes get inserted into the global list before they get fully
3473 * initialized and stay there until UMA decides to free the memory. This in
3474 * particular means the target can be found before it becomes usable and after
3475 * it becomes recycled. Picking up such vnodes is guarded with v_holdcnt set to
3476 * VHOLD_NO_SMR.
3477 *
3478 * Note: the vnode may gain more references after we transition the count 0->1.
3479 */
3480 static bool
3481 vhold_recycle_free(struct vnode *vp)
3482 {
3483 int count;
3484
3485 mtx_assert(&vnode_list_mtx, MA_OWNED);
3486
3487 count = atomic_load_int(&vp->v_holdcnt);
3488 for (;;) {
3489 if (count & VHOLD_NO_SMR) {
3490 VNASSERT((count & ~VHOLD_NO_SMR) == 0, vp,
3491 ("non-zero hold count with flags %d\n", count));
3492 return (false);
3493 }
3494 VNASSERT(count >= 0, vp, ("invalid hold count %d\n", count));
3495 if (count > 0) {
3496 return (false);
3497 }
3498 if (atomic_fcmpset_int(&vp->v_holdcnt, &count, count + 1)) {
3499 vfs_freevnodes_dec();
3500 return (true);
3501 }
3502 }
3503 }
3504
3505 static void __noinline
3506 vdbatch_process(struct vdbatch *vd)
3507 {
3508 struct vnode *vp;
3509 int i;
3510
3511 mtx_assert(&vd->lock, MA_OWNED);
3512 MPASS(curthread->td_pinned > 0);
3513 MPASS(vd->index == VDBATCH_SIZE);
3514
3515 mtx_lock(&vnode_list_mtx);
3516 critical_enter();
3517 freevnodes += vd->freevnodes;
3518 for (i = 0; i < VDBATCH_SIZE; i++) {
3519 vp = vd->tab[i];
3520 TAILQ_REMOVE(&vnode_list, vp, v_vnodelist);
3521 TAILQ_INSERT_TAIL(&vnode_list, vp, v_vnodelist);
3522 MPASS(vp->v_dbatchcpu != NOCPU);
3523 vp->v_dbatchcpu = NOCPU;
3524 }
3525 mtx_unlock(&vnode_list_mtx);
3526 vd->freevnodes = 0;
3527 bzero(vd->tab, sizeof(vd->tab));
3528 vd->index = 0;
3529 critical_exit();
3530 }
3531
3532 static void
3533 vdbatch_enqueue(struct vnode *vp)
3534 {
3535 struct vdbatch *vd;
3536
3537 ASSERT_VI_LOCKED(vp, __func__);
3538 VNASSERT(!VN_IS_DOOMED(vp), vp,
3539 ("%s: deferring requeue of a doomed vnode", __func__));
3540
3541 if (vp->v_dbatchcpu != NOCPU) {
3542 VI_UNLOCK(vp);
3543 return;
3544 }
3545
3546 sched_pin();
3547 vd = DPCPU_PTR(vd);
3548 mtx_lock(&vd->lock);
3549 MPASS(vd->index < VDBATCH_SIZE);
3550 MPASS(vd->tab[vd->index] == NULL);
3551 /*
3552 * A hack: we depend on being pinned so that we know what to put in
3553 * ->v_dbatchcpu.
3554 */
3555 vp->v_dbatchcpu = curcpu;
3556 vd->tab[vd->index] = vp;
3557 vd->index++;
3558 VI_UNLOCK(vp);
3559 if (vd->index == VDBATCH_SIZE)
3560 vdbatch_process(vd);
3561 mtx_unlock(&vd->lock);
3562 sched_unpin();
3563 }
3564
3565 /*
3566 * This routine must only be called for vnodes which are about to be
3567 * deallocated. Supporting dequeue for arbitrary vndoes would require
3568 * validating that the locked batch matches.
3569 */
3570 static void
3571 vdbatch_dequeue(struct vnode *vp)
3572 {
3573 struct vdbatch *vd;
3574 int i;
3575 short cpu;
3576
3577 VNASSERT(vp->v_type == VBAD || vp->v_type == VNON, vp,
3578 ("%s: called for a used vnode\n", __func__));
3579
3580 cpu = vp->v_dbatchcpu;
3581 if (cpu == NOCPU)
3582 return;
3583
3584 vd = DPCPU_ID_PTR(cpu, vd);
3585 mtx_lock(&vd->lock);
3586 for (i = 0; i < vd->index; i++) {
3587 if (vd->tab[i] != vp)
3588 continue;
3589 vp->v_dbatchcpu = NOCPU;
3590 vd->index--;
3591 vd->tab[i] = vd->tab[vd->index];
3592 vd->tab[vd->index] = NULL;
3593 break;
3594 }
3595 mtx_unlock(&vd->lock);
3596 /*
3597 * Either we dequeued the vnode above or the target CPU beat us to it.
3598 */
3599 MPASS(vp->v_dbatchcpu == NOCPU);
3600 }
3601
3602 /*
3603 * Drop the hold count of the vnode. If this is the last reference to
3604 * the vnode we place it on the free list unless it has been vgone'd
3605 * (marked VIRF_DOOMED) in which case we will free it.
3606 *
3607 * Because the vnode vm object keeps a hold reference on the vnode if
3608 * there is at least one resident non-cached page, the vnode cannot
3609 * leave the active list without the page cleanup done.
3610 */
3611 static void __noinline
3612 vdropl_final(struct vnode *vp)
3613 {
3614
3615 ASSERT_VI_LOCKED(vp, __func__);
3616 VNPASS(VN_IS_DOOMED(vp), vp);
3617 /*
3618 * Set the VHOLD_NO_SMR flag.
3619 *
3620 * We may be racing against vhold_smr. If they win we can just pretend
3621 * we never got this far, they will vdrop later.
3622 */
3623 if (__predict_false(!atomic_cmpset_int(&vp->v_holdcnt, 0, VHOLD_NO_SMR))) {
3624 vfs_freevnodes_inc();
3625 VI_UNLOCK(vp);
3626 /*
3627 * We lost the aforementioned race. Any subsequent access is
3628 * invalid as they might have managed to vdropl on their own.
3629 */
3630 return;
3631 }
3632 /*
3633 * Don't bump freevnodes as this one is going away.
3634 */
3635 freevnode(vp);
3636 }
3637
3638 void
3639 vdrop(struct vnode *vp)
3640 {
3641
3642 ASSERT_VI_UNLOCKED(vp, __func__);
3643 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3644 if (refcount_release_if_not_last(&vp->v_holdcnt))
3645 return;
3646 VI_LOCK(vp);
3647 vdropl(vp);
3648 }
3649
3650 static void __always_inline
3651 vdropl_impl(struct vnode *vp, bool enqueue)
3652 {
3653
3654 ASSERT_VI_LOCKED(vp, __func__);
3655 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3656 if (!refcount_release(&vp->v_holdcnt)) {
3657 VI_UNLOCK(vp);
3658 return;
3659 }
3660 VNPASS((vp->v_iflag & VI_OWEINACT) == 0, vp);
3661 VNPASS((vp->v_iflag & VI_DEFINACT) == 0, vp);
3662 if (VN_IS_DOOMED(vp)) {
3663 vdropl_final(vp);
3664 return;
3665 }
3666
3667 vfs_freevnodes_inc();
3668 if (vp->v_mflag & VMP_LAZYLIST) {
3669 vunlazy(vp);
3670 }
3671
3672 if (!enqueue) {
3673 VI_UNLOCK(vp);
3674 return;
3675 }
3676
3677 /*
3678 * Also unlocks the interlock. We can't assert on it as we
3679 * released our hold and by now the vnode might have been
3680 * freed.
3681 */
3682 vdbatch_enqueue(vp);
3683 }
3684
3685 void
3686 vdropl(struct vnode *vp)
3687 {
3688
3689 vdropl_impl(vp, true);
3690 }
3691
3692 /*
3693 * vdrop a vnode when recycling
3694 *
3695 * This is a special case routine only to be used when recycling, differs from
3696 * regular vdrop by not requeieing the vnode on LRU.
3697 *
3698 * Consider a case where vtryrecycle continuously fails with all vnodes (due to
3699 * e.g., frozen writes on the filesystem), filling the batch and causing it to
3700 * be requeued. Then vnlru will end up revisiting the same vnodes. This is a
3701 * loop which can last for as long as writes are frozen.
3702 */
3703 static void
3704 vdropl_recycle(struct vnode *vp)
3705 {
3706
3707 vdropl_impl(vp, false);
3708 }
3709
3710 static void
3711 vdrop_recycle(struct vnode *vp)
3712 {
3713
3714 VI_LOCK(vp);
3715 vdropl_recycle(vp);
3716 }
3717
3718 /*
3719 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
3720 * flags. DOINGINACT prevents us from recursing in calls to vinactive.
3721 */
3722 static int
3723 vinactivef(struct vnode *vp)
3724 {
3725 struct vm_object *obj;
3726 int error;
3727
3728 ASSERT_VOP_ELOCKED(vp, "vinactive");
3729 ASSERT_VI_LOCKED(vp, "vinactive");
3730 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
3731 ("vinactive: recursed on VI_DOINGINACT"));
3732 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3733 vp->v_iflag |= VI_DOINGINACT;
3734 vp->v_iflag &= ~VI_OWEINACT;
3735 VI_UNLOCK(vp);
3736 /*
3737 * Before moving off the active list, we must be sure that any
3738 * modified pages are converted into the vnode's dirty
3739 * buffers, since these will no longer be checked once the
3740 * vnode is on the inactive list.
3741 *
3742 * The write-out of the dirty pages is asynchronous. At the
3743 * point that VOP_INACTIVE() is called, there could still be
3744 * pending I/O and dirty pages in the object.
3745 */
3746 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
3747 vm_object_mightbedirty(obj)) {
3748 VM_OBJECT_WLOCK(obj);
3749 vm_object_page_clean(obj, 0, 0, 0);
3750 VM_OBJECT_WUNLOCK(obj);
3751 }
3752 error = VOP_INACTIVE(vp);
3753 VI_LOCK(vp);
3754 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
3755 ("vinactive: lost VI_DOINGINACT"));
3756 vp->v_iflag &= ~VI_DOINGINACT;
3757 return (error);
3758 }
3759
3760 int
3761 vinactive(struct vnode *vp)
3762 {
3763
3764 ASSERT_VOP_ELOCKED(vp, "vinactive");
3765 ASSERT_VI_LOCKED(vp, "vinactive");
3766 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3767
3768 if ((vp->v_iflag & VI_OWEINACT) == 0)
3769 return (0);
3770 if (vp->v_iflag & VI_DOINGINACT)
3771 return (0);
3772 if (vp->v_usecount > 0) {
3773 vp->v_iflag &= ~VI_OWEINACT;
3774 return (0);
3775 }
3776 return (vinactivef(vp));
3777 }
3778
3779 /*
3780 * Remove any vnodes in the vnode table belonging to mount point mp.
3781 *
3782 * If FORCECLOSE is not specified, there should not be any active ones,
3783 * return error if any are found (nb: this is a user error, not a
3784 * system error). If FORCECLOSE is specified, detach any active vnodes
3785 * that are found.
3786 *
3787 * If WRITECLOSE is set, only flush out regular file vnodes open for
3788 * writing.
3789 *
3790 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
3791 *
3792 * `rootrefs' specifies the base reference count for the root vnode
3793 * of this filesystem. The root vnode is considered busy if its
3794 * v_usecount exceeds this value. On a successful return, vflush(, td)
3795 * will call vrele() on the root vnode exactly rootrefs times.
3796 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
3797 * be zero.
3798 */
3799 #ifdef DIAGNOSTIC
3800 static int busyprt = 0; /* print out busy vnodes */
3801 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
3802 #endif
3803
3804 int
3805 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
3806 {
3807 struct vnode *vp, *mvp, *rootvp = NULL;
3808 struct vattr vattr;
3809 int busy = 0, error;
3810
3811 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
3812 rootrefs, flags);
3813 if (rootrefs > 0) {
3814 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
3815 ("vflush: bad args"));
3816 /*
3817 * Get the filesystem root vnode. We can vput() it
3818 * immediately, since with rootrefs > 0, it won't go away.
3819 */
3820 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
3821 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
3822 __func__, error);
3823 return (error);
3824 }
3825 vput(rootvp);
3826 }
3827 loop:
3828 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
3829 vholdl(vp);
3830 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
3831 if (error) {
3832 vdrop(vp);
3833 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3834 goto loop;
3835 }
3836 /*
3837 * Skip over a vnodes marked VV_SYSTEM.
3838 */
3839 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
3840 VOP_UNLOCK(vp);
3841 vdrop(vp);
3842 continue;
3843 }
3844 /*
3845 * If WRITECLOSE is set, flush out unlinked but still open
3846 * files (even if open only for reading) and regular file
3847 * vnodes open for writing.
3848 */
3849 if (flags & WRITECLOSE) {
3850 if (vp->v_object != NULL) {
3851 VM_OBJECT_WLOCK(vp->v_object);
3852 vm_object_page_clean(vp->v_object, 0, 0, 0);
3853 VM_OBJECT_WUNLOCK(vp->v_object);
3854 }
3855 do {
3856 error = VOP_FSYNC(vp, MNT_WAIT, td);
3857 } while (error == ERELOOKUP);
3858 if (error != 0) {
3859 VOP_UNLOCK(vp);
3860 vdrop(vp);
3861 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3862 return (error);
3863 }
3864 error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3865 VI_LOCK(vp);
3866
3867 if ((vp->v_type == VNON ||
3868 (error == 0 && vattr.va_nlink > 0)) &&
3869 (vp->v_writecount <= 0 || vp->v_type != VREG)) {
3870 VOP_UNLOCK(vp);
3871 vdropl(vp);
3872 continue;
3873 }
3874 } else
3875 VI_LOCK(vp);
3876 /*
3877 * With v_usecount == 0, all we need to do is clear out the
3878 * vnode data structures and we are done.
3879 *
3880 * If FORCECLOSE is set, forcibly close the vnode.
3881 */
3882 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3883 vgonel(vp);
3884 } else {
3885 busy++;
3886 #ifdef DIAGNOSTIC
3887 if (busyprt)
3888 vn_printf(vp, "vflush: busy vnode ");
3889 #endif
3890 }
3891 VOP_UNLOCK(vp);
3892 vdropl(vp);
3893 }
3894 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3895 /*
3896 * If just the root vnode is busy, and if its refcount
3897 * is equal to `rootrefs', then go ahead and kill it.
3898 */
3899 VI_LOCK(rootvp);
3900 KASSERT(busy > 0, ("vflush: not busy"));
3901 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3902 ("vflush: usecount %d < rootrefs %d",
3903 rootvp->v_usecount, rootrefs));
3904 if (busy == 1 && rootvp->v_usecount == rootrefs) {
3905 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3906 vgone(rootvp);
3907 VOP_UNLOCK(rootvp);
3908 busy = 0;
3909 } else
3910 VI_UNLOCK(rootvp);
3911 }
3912 if (busy) {
3913 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3914 busy);
3915 return (EBUSY);
3916 }
3917 for (; rootrefs > 0; rootrefs--)
3918 vrele(rootvp);
3919 return (0);
3920 }
3921
3922 /*
3923 * Recycle an unused vnode to the front of the free list.
3924 */
3925 int
3926 vrecycle(struct vnode *vp)
3927 {
3928 int recycled;
3929
3930 VI_LOCK(vp);
3931 recycled = vrecyclel(vp);
3932 VI_UNLOCK(vp);
3933 return (recycled);
3934 }
3935
3936 /*
3937 * vrecycle, with the vp interlock held.
3938 */
3939 int
3940 vrecyclel(struct vnode *vp)
3941 {
3942 int recycled;
3943
3944 ASSERT_VOP_ELOCKED(vp, __func__);
3945 ASSERT_VI_LOCKED(vp, __func__);
3946 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3947 recycled = 0;
3948 if (vp->v_usecount == 0) {
3949 recycled = 1;
3950 vgonel(vp);
3951 }
3952 return (recycled);
3953 }
3954
3955 /*
3956 * Eliminate all activity associated with a vnode
3957 * in preparation for reuse.
3958 */
3959 void
3960 vgone(struct vnode *vp)
3961 {
3962 VI_LOCK(vp);
3963 vgonel(vp);
3964 VI_UNLOCK(vp);
3965 }
3966
3967 /*
3968 * Notify upper mounts about reclaimed or unlinked vnode.
3969 */
3970 void
3971 vfs_notify_upper(struct vnode *vp, enum vfs_notify_upper_type event)
3972 {
3973 struct mount *mp;
3974 struct mount_upper_node *ump;
3975
3976 mp = atomic_load_ptr(&vp->v_mount);
3977 if (mp == NULL)
3978 return;
3979 if (TAILQ_EMPTY(&mp->mnt_notify))
3980 return;
3981
3982 MNT_ILOCK(mp);
3983 mp->mnt_upper_pending++;
3984 KASSERT(mp->mnt_upper_pending > 0,
3985 ("%s: mnt_upper_pending %d", __func__, mp->mnt_upper_pending));
3986 TAILQ_FOREACH(ump, &mp->mnt_notify, mnt_upper_link) {
3987 MNT_IUNLOCK(mp);
3988 switch (event) {
3989 case VFS_NOTIFY_UPPER_RECLAIM:
3990 VFS_RECLAIM_LOWERVP(ump->mp, vp);
3991 break;
3992 case VFS_NOTIFY_UPPER_UNLINK:
3993 VFS_UNLINK_LOWERVP(ump->mp, vp);
3994 break;
3995 }
3996 MNT_ILOCK(mp);
3997 }
3998 mp->mnt_upper_pending--;
3999 if ((mp->mnt_kern_flag & MNTK_UPPER_WAITER) != 0 &&
4000 mp->mnt_upper_pending == 0) {
4001 mp->mnt_kern_flag &= ~MNTK_UPPER_WAITER;
4002 wakeup(&mp->mnt_uppers);
4003 }
4004 MNT_IUNLOCK(mp);
4005 }
4006
4007 /*
4008 * vgone, with the vp interlock held.
4009 */
4010 static void
4011 vgonel(struct vnode *vp)
4012 {
4013 struct thread *td;
4014 struct mount *mp;
4015 vm_object_t object;
4016 bool active, doinginact, oweinact;
4017
4018 ASSERT_VOP_ELOCKED(vp, "vgonel");
4019 ASSERT_VI_LOCKED(vp, "vgonel");
4020 VNASSERT(vp->v_holdcnt, vp,
4021 ("vgonel: vp %p has no reference.", vp));
4022 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
4023 td = curthread;
4024
4025 /*
4026 * Don't vgonel if we're already doomed.
4027 */
4028 if (VN_IS_DOOMED(vp)) {
4029 VNPASS(vn_get_state(vp) == VSTATE_DESTROYING || \
4030 vn_get_state(vp) == VSTATE_DEAD, vp);
4031 return;
4032 }
4033 /*
4034 * Paired with freevnode.
4035 */
4036 vn_seqc_write_begin_locked(vp);
4037 vunlazy_gone(vp);
4038 vn_irflag_set_locked(vp, VIRF_DOOMED);
4039 vn_set_state(vp, VSTATE_DESTROYING);
4040
4041 /*
4042 * Check to see if the vnode is in use. If so, we have to
4043 * call VOP_CLOSE() and VOP_INACTIVE().
4044 *
4045 * It could be that VOP_INACTIVE() requested reclamation, in
4046 * which case we should avoid recursion, so check
4047 * VI_DOINGINACT. This is not precise but good enough.
4048 */
4049 active = vp->v_usecount > 0;
4050 oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4051 doinginact = (vp->v_iflag & VI_DOINGINACT) != 0;
4052
4053 /*
4054 * If we need to do inactive VI_OWEINACT will be set.
4055 */
4056 if (vp->v_iflag & VI_DEFINACT) {
4057 VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
4058 vp->v_iflag &= ~VI_DEFINACT;
4059 vdropl(vp);
4060 } else {
4061 VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
4062 VI_UNLOCK(vp);
4063 }
4064 cache_purge_vgone(vp);
4065 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
4066
4067 /*
4068 * If purging an active vnode, it must be closed and
4069 * deactivated before being reclaimed.
4070 */
4071 if (active)
4072 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
4073 if (!doinginact) {
4074 do {
4075 if (oweinact || active) {
4076 VI_LOCK(vp);
4077 vinactivef(vp);
4078 oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
4079 VI_UNLOCK(vp);
4080 }
4081 } while (oweinact);
4082 }
4083 if (vp->v_type == VSOCK)
4084 vfs_unp_reclaim(vp);
4085
4086 /*
4087 * Clean out any buffers associated with the vnode.
4088 * If the flush fails, just toss the buffers.
4089 */
4090 mp = NULL;
4091 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
4092 (void) vn_start_secondary_write(vp, &mp, V_WAIT);
4093 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
4094 while (vinvalbuf(vp, 0, 0, 0) != 0)
4095 ;
4096 }
4097
4098 BO_LOCK(&vp->v_bufobj);
4099 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
4100 vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
4101 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
4102 vp->v_bufobj.bo_clean.bv_cnt == 0,
4103 ("vp %p bufobj not invalidated", vp));
4104
4105 /*
4106 * For VMIO bufobj, BO_DEAD is set later, or in
4107 * vm_object_terminate() after the object's page queue is
4108 * flushed.
4109 */
4110 object = vp->v_bufobj.bo_object;
4111 if (object == NULL)
4112 vp->v_bufobj.bo_flag |= BO_DEAD;
4113 BO_UNLOCK(&vp->v_bufobj);
4114
4115 /*
4116 * Handle the VM part. Tmpfs handles v_object on its own (the
4117 * OBJT_VNODE check). Nullfs or other bypassing filesystems
4118 * should not touch the object borrowed from the lower vnode
4119 * (the handle check).
4120 */
4121 if (object != NULL && object->type == OBJT_VNODE &&
4122 object->handle == vp)
4123 vnode_destroy_vobject(vp);
4124
4125 /*
4126 * Reclaim the vnode.
4127 */
4128 if (VOP_RECLAIM(vp))
4129 panic("vgone: cannot reclaim");
4130 if (mp != NULL)
4131 vn_finished_secondary_write(mp);
4132 VNASSERT(vp->v_object == NULL, vp,
4133 ("vop_reclaim left v_object vp=%p", vp));
4134 /*
4135 * Clear the advisory locks and wake up waiting threads.
4136 */
4137 if (vp->v_lockf != NULL) {
4138 (void)VOP_ADVLOCKPURGE(vp);
4139 vp->v_lockf = NULL;
4140 }
4141 /*
4142 * Delete from old mount point vnode list.
4143 */
4144 if (vp->v_mount == NULL) {
4145 VI_LOCK(vp);
4146 } else {
4147 delmntque(vp);
4148 ASSERT_VI_LOCKED(vp, "vgonel 2");
4149 }
4150 /*
4151 * Done with purge, reset to the standard lock and invalidate
4152 * the vnode.
4153 */
4154 vp->v_vnlock = &vp->v_lock;
4155 vp->v_op = &dead_vnodeops;
4156 vp->v_type = VBAD;
4157 vn_set_state(vp, VSTATE_DEAD);
4158 }
4159
4160 /*
4161 * Print out a description of a vnode.
4162 */
4163 static const char *const vtypename[] = {
4164 [VNON] = "VNON",
4165 [VREG] = "VREG",
4166 [VDIR] = "VDIR",
4167 [VBLK] = "VBLK",
4168 [VCHR] = "VCHR",
4169 [VLNK] = "VLNK",
4170 [VSOCK] = "VSOCK",
4171 [VFIFO] = "VFIFO",
4172 [VBAD] = "VBAD",
4173 [VMARKER] = "VMARKER",
4174 };
4175 _Static_assert(nitems(vtypename) == VLASTTYPE + 1,
4176 "vnode type name not added to vtypename");
4177
4178 static const char *const vstatename[] = {
4179 [VSTATE_UNINITIALIZED] = "VSTATE_UNINITIALIZED",
4180 [VSTATE_CONSTRUCTED] = "VSTATE_CONSTRUCTED",
4181 [VSTATE_DESTROYING] = "VSTATE_DESTROYING",
4182 [VSTATE_DEAD] = "VSTATE_DEAD",
4183 };
4184 _Static_assert(nitems(vstatename) == VLASTSTATE + 1,
4185 "vnode state name not added to vstatename");
4186
4187 _Static_assert((VHOLD_ALL_FLAGS & ~VHOLD_NO_SMR) == 0,
4188 "new hold count flag not added to vn_printf");
4189
4190 void
4191 vn_printf(struct vnode *vp, const char *fmt, ...)
4192 {
4193 va_list ap;
4194 char buf[256], buf2[16];
4195 u_long flags;
4196 u_int holdcnt;
4197 short irflag;
4198
4199 va_start(ap, fmt);
4200 vprintf(fmt, ap);
4201 va_end(ap);
4202 printf("%p: ", (void *)vp);
4203 printf("type %s state %s\n", vtypename[vp->v_type], vstatename[vp->v_state]);
4204 holdcnt = atomic_load_int(&vp->v_holdcnt);
4205 printf(" usecount %d, writecount %d, refcount %d seqc users %d",
4206 vp->v_usecount, vp->v_writecount, holdcnt & ~VHOLD_ALL_FLAGS,
4207 vp->v_seqc_users);
4208 switch (vp->v_type) {
4209 case VDIR:
4210 printf(" mountedhere %p\n", vp->v_mountedhere);
4211 break;
4212 case VCHR:
4213 printf(" rdev %p\n", vp->v_rdev);
4214 break;
4215 case VSOCK:
4216 printf(" socket %p\n", vp->v_unpcb);
4217 break;
4218 case VFIFO:
4219 printf(" fifoinfo %p\n", vp->v_fifoinfo);
4220 break;
4221 default:
4222 printf("\n");
4223 break;
4224 }
4225 buf[0] = '\0';
4226 buf[1] = '\0';
4227 if (holdcnt & VHOLD_NO_SMR)
4228 strlcat(buf, "|VHOLD_NO_SMR", sizeof(buf));
4229 printf(" hold count flags (%s)\n", buf + 1);
4230
4231 buf[0] = '\0';
4232 buf[1] = '\0';
4233 irflag = vn_irflag_read(vp);
4234 if (irflag & VIRF_DOOMED)
4235 strlcat(buf, "|VIRF_DOOMED", sizeof(buf));
4236 if (irflag & VIRF_PGREAD)
4237 strlcat(buf, "|VIRF_PGREAD", sizeof(buf));
4238 if (irflag & VIRF_MOUNTPOINT)
4239 strlcat(buf, "|VIRF_MOUNTPOINT", sizeof(buf));
4240 if (irflag & VIRF_TEXT_REF)
4241 strlcat(buf, "|VIRF_TEXT_REF", sizeof(buf));
4242 flags = irflag & ~(VIRF_DOOMED | VIRF_PGREAD | VIRF_MOUNTPOINT | VIRF_TEXT_REF);
4243 if (flags != 0) {
4244 snprintf(buf2, sizeof(buf2), "|VIRF(0x%lx)", flags);
4245 strlcat(buf, buf2, sizeof(buf));
4246 }
4247 if (vp->v_vflag & VV_ROOT)
4248 strlcat(buf, "|VV_ROOT", sizeof(buf));
4249 if (vp->v_vflag & VV_ISTTY)
4250 strlcat(buf, "|VV_ISTTY", sizeof(buf));
4251 if (vp->v_vflag & VV_NOSYNC)
4252 strlcat(buf, "|VV_NOSYNC", sizeof(buf));
4253 if (vp->v_vflag & VV_ETERNALDEV)
4254 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
4255 if (vp->v_vflag & VV_CACHEDLABEL)
4256 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
4257 if (vp->v_vflag & VV_VMSIZEVNLOCK)
4258 strlcat(buf, "|VV_VMSIZEVNLOCK", sizeof(buf));
4259 if (vp->v_vflag & VV_COPYONWRITE)
4260 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
4261 if (vp->v_vflag & VV_SYSTEM)
4262 strlcat(buf, "|VV_SYSTEM", sizeof(buf));
4263 if (vp->v_vflag & VV_PROCDEP)
4264 strlcat(buf, "|VV_PROCDEP", sizeof(buf));
4265 if (vp->v_vflag & VV_DELETED)
4266 strlcat(buf, "|VV_DELETED", sizeof(buf));
4267 if (vp->v_vflag & VV_MD)
4268 strlcat(buf, "|VV_MD", sizeof(buf));
4269 if (vp->v_vflag & VV_FORCEINSMQ)
4270 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
4271 if (vp->v_vflag & VV_READLINK)
4272 strlcat(buf, "|VV_READLINK", sizeof(buf));
4273 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
4274 VV_CACHEDLABEL | VV_VMSIZEVNLOCK | VV_COPYONWRITE | VV_SYSTEM |
4275 VV_PROCDEP | VV_DELETED | VV_MD | VV_FORCEINSMQ | VV_READLINK);
4276 if (flags != 0) {
4277 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
4278 strlcat(buf, buf2, sizeof(buf));
4279 }
4280 if (vp->v_iflag & VI_MOUNT)
4281 strlcat(buf, "|VI_MOUNT", sizeof(buf));
4282 if (vp->v_iflag & VI_DOINGINACT)
4283 strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
4284 if (vp->v_iflag & VI_OWEINACT)
4285 strlcat(buf, "|VI_OWEINACT", sizeof(buf));
4286 if (vp->v_iflag & VI_DEFINACT)
4287 strlcat(buf, "|VI_DEFINACT", sizeof(buf));
4288 if (vp->v_iflag & VI_FOPENING)
4289 strlcat(buf, "|VI_FOPENING", sizeof(buf));
4290 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOINGINACT |
4291 VI_OWEINACT | VI_DEFINACT | VI_FOPENING);
4292 if (flags != 0) {
4293 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
4294 strlcat(buf, buf2, sizeof(buf));
4295 }
4296 if (vp->v_mflag & VMP_LAZYLIST)
4297 strlcat(buf, "|VMP_LAZYLIST", sizeof(buf));
4298 flags = vp->v_mflag & ~(VMP_LAZYLIST);
4299 if (flags != 0) {
4300 snprintf(buf2, sizeof(buf2), "|VMP(0x%lx)", flags);
4301 strlcat(buf, buf2, sizeof(buf));
4302 }
4303 printf(" flags (%s)", buf + 1);
4304 if (mtx_owned(VI_MTX(vp)))
4305 printf(" VI_LOCKed");
4306 printf("\n");
4307 if (vp->v_object != NULL)
4308 printf(" v_object %p ref %d pages %d "
4309 "cleanbuf %d dirtybuf %d\n",
4310 vp->v_object, vp->v_object->ref_count,
4311 vp->v_object->resident_page_count,
4312 vp->v_bufobj.bo_clean.bv_cnt,
4313 vp->v_bufobj.bo_dirty.bv_cnt);
4314 printf(" ");
4315 lockmgr_printinfo(vp->v_vnlock);
4316 if (vp->v_data != NULL)
4317 VOP_PRINT(vp);
4318 }
4319
4320 #ifdef DDB
4321 /*
4322 * List all of the locked vnodes in the system.
4323 * Called when debugging the kernel.
4324 */
4325 DB_SHOW_COMMAND_FLAGS(lockedvnods, lockedvnodes, DB_CMD_MEMSAFE)
4326 {
4327 struct mount *mp;
4328 struct vnode *vp;
4329
4330 /*
4331 * Note: because this is DDB, we can't obey the locking semantics
4332 * for these structures, which means we could catch an inconsistent
4333 * state and dereference a nasty pointer. Not much to be done
4334 * about that.
4335 */
4336 db_printf("Locked vnodes\n");
4337 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4338 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4339 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
4340 vn_printf(vp, "vnode ");
4341 }
4342 }
4343 }
4344
4345 /*
4346 * Show details about the given vnode.
4347 */
4348 DB_SHOW_COMMAND(vnode, db_show_vnode)
4349 {
4350 struct vnode *vp;
4351
4352 if (!have_addr)
4353 return;
4354 vp = (struct vnode *)addr;
4355 vn_printf(vp, "vnode ");
4356 }
4357
4358 /*
4359 * Show details about the given mount point.
4360 */
4361 DB_SHOW_COMMAND(mount, db_show_mount)
4362 {
4363 struct mount *mp;
4364 struct vfsopt *opt;
4365 struct statfs *sp;
4366 struct vnode *vp;
4367 char buf[512];
4368 uint64_t mflags;
4369 u_int flags;
4370
4371 if (!have_addr) {
4372 /* No address given, print short info about all mount points. */
4373 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4374 db_printf("%p %s on %s (%s)\n", mp,
4375 mp->mnt_stat.f_mntfromname,
4376 mp->mnt_stat.f_mntonname,
4377 mp->mnt_stat.f_fstypename);
4378 if (db_pager_quit)
4379 break;
4380 }
4381 db_printf("\nMore info: show mount <addr>\n");
4382 return;
4383 }
4384
4385 mp = (struct mount *)addr;
4386 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
4387 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
4388
4389 buf[0] = '\0';
4390 mflags = mp->mnt_flag;
4391 #define MNT_FLAG(flag) do { \
4392 if (mflags & (flag)) { \
4393 if (buf[0] != '\0') \
4394 strlcat(buf, ", ", sizeof(buf)); \
4395 strlcat(buf, (#flag) + 4, sizeof(buf)); \
4396 mflags &= ~(flag); \
4397 } \
4398 } while (0)
4399 MNT_FLAG(MNT_RDONLY);
4400 MNT_FLAG(MNT_SYNCHRONOUS);
4401 MNT_FLAG(MNT_NOEXEC);
4402 MNT_FLAG(MNT_NOSUID);
4403 MNT_FLAG(MNT_NFS4ACLS);
4404 MNT_FLAG(MNT_UNION);
4405 MNT_FLAG(MNT_ASYNC);
4406 MNT_FLAG(MNT_SUIDDIR);
4407 MNT_FLAG(MNT_SOFTDEP);
4408 MNT_FLAG(MNT_NOSYMFOLLOW);
4409 MNT_FLAG(MNT_GJOURNAL);
4410 MNT_FLAG(MNT_MULTILABEL);
4411 MNT_FLAG(MNT_ACLS);
4412 MNT_FLAG(MNT_NOATIME);
4413 MNT_FLAG(MNT_NOCLUSTERR);
4414 MNT_FLAG(MNT_NOCLUSTERW);
4415 MNT_FLAG(MNT_SUJ);
4416 MNT_FLAG(MNT_EXRDONLY);
4417 MNT_FLAG(MNT_EXPORTED);
4418 MNT_FLAG(MNT_DEFEXPORTED);
4419 MNT_FLAG(MNT_EXPORTANON);
4420 MNT_FLAG(MNT_EXKERB);
4421 MNT_FLAG(MNT_EXPUBLIC);
4422 MNT_FLAG(MNT_LOCAL);
4423 MNT_FLAG(MNT_QUOTA);
4424 MNT_FLAG(MNT_ROOTFS);
4425 MNT_FLAG(MNT_USER);
4426 MNT_FLAG(MNT_IGNORE);
4427 MNT_FLAG(MNT_UPDATE);
4428 MNT_FLAG(MNT_DELEXPORT);
4429 MNT_FLAG(MNT_RELOAD);
4430 MNT_FLAG(MNT_FORCE);
4431 MNT_FLAG(MNT_SNAPSHOT);
4432 MNT_FLAG(MNT_BYFSID);
4433 #undef MNT_FLAG
4434 if (mflags != 0) {
4435 if (buf[0] != '\0')
4436 strlcat(buf, ", ", sizeof(buf));
4437 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4438 "0x%016jx", mflags);
4439 }
4440 db_printf(" mnt_flag = %s\n", buf);
4441
4442 buf[0] = '\0';
4443 flags = mp->mnt_kern_flag;
4444 #define MNT_KERN_FLAG(flag) do { \
4445 if (flags & (flag)) { \
4446 if (buf[0] != '\0') \
4447 strlcat(buf, ", ", sizeof(buf)); \
4448 strlcat(buf, (#flag) + 5, sizeof(buf)); \
4449 flags &= ~(flag); \
4450 } \
4451 } while (0)
4452 MNT_KERN_FLAG(MNTK_UNMOUNTF);
4453 MNT_KERN_FLAG(MNTK_ASYNC);
4454 MNT_KERN_FLAG(MNTK_SOFTDEP);
4455 MNT_KERN_FLAG(MNTK_NOMSYNC);
4456 MNT_KERN_FLAG(MNTK_DRAINING);
4457 MNT_KERN_FLAG(MNTK_REFEXPIRE);
4458 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
4459 MNT_KERN_FLAG(MNTK_SHARED_WRITES);
4460 MNT_KERN_FLAG(MNTK_NO_IOPF);
4461 MNT_KERN_FLAG(MNTK_RECURSE);
4462 MNT_KERN_FLAG(MNTK_UPPER_WAITER);
4463 MNT_KERN_FLAG(MNTK_UNLOCKED_INSMNTQUE);
4464 MNT_KERN_FLAG(MNTK_USES_BCACHE);
4465 MNT_KERN_FLAG(MNTK_VMSETSIZE_BUG);
4466 MNT_KERN_FLAG(MNTK_FPLOOKUP);
4467 MNT_KERN_FLAG(MNTK_TASKQUEUE_WAITER);
4468 MNT_KERN_FLAG(MNTK_NOASYNC);
4469 MNT_KERN_FLAG(MNTK_UNMOUNT);
4470 MNT_KERN_FLAG(MNTK_MWAIT);
4471 MNT_KERN_FLAG(MNTK_SUSPEND);
4472 MNT_KERN_FLAG(MNTK_SUSPEND2);
4473 MNT_KERN_FLAG(MNTK_SUSPENDED);
4474 MNT_KERN_FLAG(MNTK_NULL_NOCACHE);
4475 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
4476 #undef MNT_KERN_FLAG
4477 if (flags != 0) {
4478 if (buf[0] != '\0')
4479 strlcat(buf, ", ", sizeof(buf));
4480 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
4481 "0x%08x", flags);
4482 }
4483 db_printf(" mnt_kern_flag = %s\n", buf);
4484
4485 db_printf(" mnt_opt = ");
4486 opt = TAILQ_FIRST(mp->mnt_opt);
4487 if (opt != NULL) {
4488 db_printf("%s", opt->name);
4489 opt = TAILQ_NEXT(opt, link);
4490 while (opt != NULL) {
4491 db_printf(", %s", opt->name);
4492 opt = TAILQ_NEXT(opt, link);
4493 }
4494 }
4495 db_printf("\n");
4496
4497 sp = &mp->mnt_stat;
4498 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx "
4499 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
4500 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
4501 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
4502 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
4503 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
4504 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
4505 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
4506 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
4507 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
4508 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
4509 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
4510
4511 db_printf(" mnt_cred = { uid=%u ruid=%u",
4512 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
4513 if (jailed(mp->mnt_cred))
4514 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
4515 db_printf(" }\n");
4516 db_printf(" mnt_ref = %d (with %d in the struct)\n",
4517 vfs_mount_fetch_counter(mp, MNT_COUNT_REF), mp->mnt_ref);
4518 db_printf(" mnt_gen = %d\n", mp->mnt_gen);
4519 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
4520 db_printf(" mnt_lazyvnodelistsize = %d\n",
4521 mp->mnt_lazyvnodelistsize);
4522 db_printf(" mnt_writeopcount = %d (with %d in the struct)\n",
4523 vfs_mount_fetch_counter(mp, MNT_COUNT_WRITEOPCOUNT), mp->mnt_writeopcount);
4524 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max);
4525 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed);
4526 db_printf(" mnt_lockref = %d (with %d in the struct)\n",
4527 vfs_mount_fetch_counter(mp, MNT_COUNT_LOCKREF), mp->mnt_lockref);
4528 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
4529 db_printf(" mnt_secondary_accwrites = %d\n",
4530 mp->mnt_secondary_accwrites);
4531 db_printf(" mnt_gjprovider = %s\n",
4532 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
4533 db_printf(" mnt_vfs_ops = %d\n", mp->mnt_vfs_ops);
4534
4535 db_printf("\n\nList of active vnodes\n");
4536 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4537 if (vp->v_type != VMARKER && vp->v_holdcnt > 0) {
4538 vn_printf(vp, "vnode ");
4539 if (db_pager_quit)
4540 break;
4541 }
4542 }
4543 db_printf("\n\nList of inactive vnodes\n");
4544 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4545 if (vp->v_type != VMARKER && vp->v_holdcnt == 0) {
4546 vn_printf(vp, "vnode ");
4547 if (db_pager_quit)
4548 break;
4549 }
4550 }
4551 }
4552 #endif /* DDB */
4553
4554 /*
4555 * Fill in a struct xvfsconf based on a struct vfsconf.
4556 */
4557 static int
4558 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
4559 {
4560 struct xvfsconf xvfsp;
4561
4562 bzero(&xvfsp, sizeof(xvfsp));
4563 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4564 xvfsp.vfc_typenum = vfsp->vfc_typenum;
4565 xvfsp.vfc_refcount = vfsp->vfc_refcount;
4566 xvfsp.vfc_flags = vfsp->vfc_flags;
4567 /*
4568 * These are unused in userland, we keep them
4569 * to not break binary compatibility.
4570 */
4571 xvfsp.vfc_vfsops = NULL;
4572 xvfsp.vfc_next = NULL;
4573 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4574 }
4575
4576 #ifdef COMPAT_FREEBSD32
4577 struct xvfsconf32 {
4578 uint32_t vfc_vfsops;
4579 char vfc_name[MFSNAMELEN];
4580 int32_t vfc_typenum;
4581 int32_t vfc_refcount;
4582 int32_t vfc_flags;
4583 uint32_t vfc_next;
4584 };
4585
4586 static int
4587 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
4588 {
4589 struct xvfsconf32 xvfsp;
4590
4591 bzero(&xvfsp, sizeof(xvfsp));
4592 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
4593 xvfsp.vfc_typenum = vfsp->vfc_typenum;
4594 xvfsp.vfc_refcount = vfsp->vfc_refcount;
4595 xvfsp.vfc_flags = vfsp->vfc_flags;
4596 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
4597 }
4598 #endif
4599
4600 /*
4601 * Top level filesystem related information gathering.
4602 */
4603 static int
4604 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
4605 {
4606 struct vfsconf *vfsp;
4607 int error;
4608
4609 error = 0;
4610 vfsconf_slock();
4611 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4612 #ifdef COMPAT_FREEBSD32
4613 if (req->flags & SCTL_MASK32)
4614 error = vfsconf2x32(req, vfsp);
4615 else
4616 #endif
4617 error = vfsconf2x(req, vfsp);
4618 if (error)
4619 break;
4620 }
4621 vfsconf_sunlock();
4622 return (error);
4623 }
4624
4625 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
4626 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
4627 "S,xvfsconf", "List of all configured filesystems");
4628
4629 #ifndef BURN_BRIDGES
4630 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
4631
4632 static int
4633 vfs_sysctl(SYSCTL_HANDLER_ARGS)
4634 {
4635 int *name = (int *)arg1 - 1; /* XXX */
4636 u_int namelen = arg2 + 1; /* XXX */
4637 struct vfsconf *vfsp;
4638
4639 log(LOG_WARNING, "userland calling deprecated sysctl, "
4640 "please rebuild world\n");
4641
4642 #if 1 || defined(COMPAT_PRELITE2)
4643 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
4644 if (namelen == 1)
4645 return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
4646 #endif
4647
4648 switch (name[1]) {
4649 case VFS_MAXTYPENUM:
4650 if (namelen != 2)
4651 return (ENOTDIR);
4652 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
4653 case VFS_CONF:
4654 if (namelen != 3)
4655 return (ENOTDIR); /* overloaded */
4656 vfsconf_slock();
4657 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4658 if (vfsp->vfc_typenum == name[2])
4659 break;
4660 }
4661 vfsconf_sunlock();
4662 if (vfsp == NULL)
4663 return (EOPNOTSUPP);
4664 #ifdef COMPAT_FREEBSD32
4665 if (req->flags & SCTL_MASK32)
4666 return (vfsconf2x32(req, vfsp));
4667 else
4668 #endif
4669 return (vfsconf2x(req, vfsp));
4670 }
4671 return (EOPNOTSUPP);
4672 }
4673
4674 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
4675 CTLFLAG_MPSAFE, vfs_sysctl,
4676 "Generic filesystem");
4677
4678 #if 1 || defined(COMPAT_PRELITE2)
4679
4680 static int
4681 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
4682 {
4683 int error;
4684 struct vfsconf *vfsp;
4685 struct ovfsconf ovfs;
4686
4687 vfsconf_slock();
4688 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
4689 bzero(&ovfs, sizeof(ovfs));
4690 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */
4691 strcpy(ovfs.vfc_name, vfsp->vfc_name);
4692 ovfs.vfc_index = vfsp->vfc_typenum;
4693 ovfs.vfc_refcount = vfsp->vfc_refcount;
4694 ovfs.vfc_flags = vfsp->vfc_flags;
4695 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
4696 if (error != 0) {
4697 vfsconf_sunlock();
4698 return (error);
4699 }
4700 }
4701 vfsconf_sunlock();
4702 return (0);
4703 }
4704
4705 #endif /* 1 || COMPAT_PRELITE2 */
4706 #endif /* !BURN_BRIDGES */
4707
4708 #define KINFO_VNODESLOP 10
4709 #ifdef notyet
4710 /*
4711 * Dump vnode list (via sysctl).
4712 */
4713 /* ARGSUSED */
4714 static int
4715 sysctl_vnode(SYSCTL_HANDLER_ARGS)
4716 {
4717 struct xvnode *xvn;
4718 struct mount *mp;
4719 struct vnode *vp;
4720 int error, len, n;
4721
4722 /*
4723 * Stale numvnodes access is not fatal here.
4724 */
4725 req->lock = 0;
4726 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
4727 if (!req->oldptr)
4728 /* Make an estimate */
4729 return (SYSCTL_OUT(req, 0, len));
4730
4731 error = sysctl_wire_old_buffer(req, 0);
4732 if (error != 0)
4733 return (error);
4734 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
4735 n = 0;
4736 mtx_lock(&mountlist_mtx);
4737 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
4738 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
4739 continue;
4740 MNT_ILOCK(mp);
4741 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
4742 if (n == len)
4743 break;
4744 vref(vp);
4745 xvn[n].xv_size = sizeof *xvn;
4746 xvn[n].xv_vnode = vp;
4747 xvn[n].xv_id = 0; /* XXX compat */
4748 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
4749 XV_COPY(usecount);
4750 XV_COPY(writecount);
4751 XV_COPY(holdcnt);
4752 XV_COPY(mount);
4753 XV_COPY(numoutput);
4754 XV_COPY(type);
4755 #undef XV_COPY
4756 xvn[n].xv_flag = vp->v_vflag;
4757
4758 switch (vp->v_type) {
4759 case VREG:
4760 case VDIR:
4761 case VLNK:
4762 break;
4763 case VBLK:
4764 case VCHR:
4765 if (vp->v_rdev == NULL) {
4766 vrele(vp);
4767 continue;
4768 }
4769 xvn[n].xv_dev = dev2udev(vp->v_rdev);
4770 break;
4771 case VSOCK:
4772 xvn[n].xv_socket = vp->v_socket;
4773 break;
4774 case VFIFO:
4775 xvn[n].xv_fifo = vp->v_fifoinfo;
4776 break;
4777 case VNON:
4778 case VBAD:
4779 default:
4780 /* shouldn't happen? */
4781 vrele(vp);
4782 continue;
4783 }
4784 vrele(vp);
4785 ++n;
4786 }
4787 MNT_IUNLOCK(mp);
4788 mtx_lock(&mountlist_mtx);
4789 vfs_unbusy(mp);
4790 if (n == len)
4791 break;
4792 }
4793 mtx_unlock(&mountlist_mtx);
4794
4795 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
4796 free(xvn, M_TEMP);
4797 return (error);
4798 }
4799
4800 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
4801 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
4802 "");
4803 #endif
4804
4805 static void
4806 unmount_or_warn(struct mount *mp)
4807 {
4808 int error;
4809
4810 error = dounmount(mp, MNT_FORCE, curthread);
4811 if (error != 0) {
4812 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
4813 if (error == EBUSY)
4814 printf("BUSY)\n");
4815 else
4816 printf("%d)\n", error);
4817 }
4818 }
4819
4820 /*
4821 * Unmount all filesystems. The list is traversed in reverse order
4822 * of mounting to avoid dependencies.
4823 */
4824 void
4825 vfs_unmountall(void)
4826 {
4827 struct mount *mp, *tmp;
4828
4829 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
4830
4831 /*
4832 * Since this only runs when rebooting, it is not interlocked.
4833 */
4834 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
4835 vfs_ref(mp);
4836
4837 /*
4838 * Forcibly unmounting "/dev" before "/" would prevent clean
4839 * unmount of the latter.
4840 */
4841 if (mp == rootdevmp)
4842 continue;
4843
4844 unmount_or_warn(mp);
4845 }
4846
4847 if (rootdevmp != NULL)
4848 unmount_or_warn(rootdevmp);
4849 }
4850
4851 static void
4852 vfs_deferred_inactive(struct vnode *vp, int lkflags)
4853 {
4854
4855 ASSERT_VI_LOCKED(vp, __func__);
4856 VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
4857 if ((vp->v_iflag & VI_OWEINACT) == 0) {
4858 vdropl(vp);
4859 return;
4860 }
4861 if (vn_lock(vp, lkflags) == 0) {
4862 VI_LOCK(vp);
4863 vinactive(vp);
4864 VOP_UNLOCK(vp);
4865 vdropl(vp);
4866 return;
4867 }
4868 vdefer_inactive_unlocked(vp);
4869 }
4870
4871 static int
4872 vfs_periodic_inactive_filter(struct vnode *vp, void *arg)
4873 {
4874
4875 return (vp->v_iflag & VI_DEFINACT);
4876 }
4877
4878 static void __noinline
4879 vfs_periodic_inactive(struct mount *mp, int flags)
4880 {
4881 struct vnode *vp, *mvp;
4882 int lkflags;
4883
4884 lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4885 if (flags != MNT_WAIT)
4886 lkflags |= LK_NOWAIT;
4887
4888 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_inactive_filter, NULL) {
4889 if ((vp->v_iflag & VI_DEFINACT) == 0) {
4890 VI_UNLOCK(vp);
4891 continue;
4892 }
4893 vp->v_iflag &= ~VI_DEFINACT;
4894 vfs_deferred_inactive(vp, lkflags);
4895 }
4896 }
4897
4898 static inline bool
4899 vfs_want_msync(struct vnode *vp)
4900 {
4901 struct vm_object *obj;
4902
4903 /*
4904 * This test may be performed without any locks held.
4905 * We rely on vm_object's type stability.
4906 */
4907 if (vp->v_vflag & VV_NOSYNC)
4908 return (false);
4909 obj = vp->v_object;
4910 return (obj != NULL && vm_object_mightbedirty(obj));
4911 }
4912
4913 static int
4914 vfs_periodic_msync_inactive_filter(struct vnode *vp, void *arg __unused)
4915 {
4916
4917 if (vp->v_vflag & VV_NOSYNC)
4918 return (false);
4919 if (vp->v_iflag & VI_DEFINACT)
4920 return (true);
4921 return (vfs_want_msync(vp));
4922 }
4923
4924 static void __noinline
4925 vfs_periodic_msync_inactive(struct mount *mp, int flags)
4926 {
4927 struct vnode *vp, *mvp;
4928 struct vm_object *obj;
4929 int lkflags, objflags;
4930 bool seen_defer;
4931
4932 lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
4933 if (flags != MNT_WAIT) {
4934 lkflags |= LK_NOWAIT;
4935 objflags = OBJPC_NOSYNC;
4936 } else {
4937 objflags = OBJPC_SYNC;
4938 }
4939
4940 MNT_VNODE_FOREACH_LAZY(vp, mp, mvp, vfs_periodic_msync_inactive_filter, NULL) {
4941 seen_defer = false;
4942 if (vp->v_iflag & VI_DEFINACT) {
4943 vp->v_iflag &= ~VI_DEFINACT;
4944 seen_defer = true;
4945 }
4946 if (!vfs_want_msync(vp)) {
4947 if (seen_defer)
4948 vfs_deferred_inactive(vp, lkflags);
4949 else
4950 VI_UNLOCK(vp);
4951 continue;
4952 }
4953 if (vget(vp, lkflags) == 0) {
4954 obj = vp->v_object;
4955 if (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0) {
4956 VM_OBJECT_WLOCK(obj);
4957 vm_object_page_clean(obj, 0, 0, objflags);
4958 VM_OBJECT_WUNLOCK(obj);
4959 }
4960 vput(vp);
4961 if (seen_defer)
4962 vdrop(vp);
4963 } else {
4964 if (seen_defer)
4965 vdefer_inactive_unlocked(vp);
4966 }
4967 }
4968 }
4969
4970 void
4971 vfs_periodic(struct mount *mp, int flags)
4972 {
4973
4974 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
4975
4976 if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
4977 vfs_periodic_inactive(mp, flags);
4978 else
4979 vfs_periodic_msync_inactive(mp, flags);
4980 }
4981
4982 static void
4983 destroy_vpollinfo_free(struct vpollinfo *vi)
4984 {
4985
4986 knlist_destroy(&vi->vpi_selinfo.si_note);
4987 mtx_destroy(&vi->vpi_lock);
4988 free(vi, M_VNODEPOLL);
4989 }
4990
4991 static void
4992 destroy_vpollinfo(struct vpollinfo *vi)
4993 {
4994
4995 knlist_clear(&vi->vpi_selinfo.si_note, 1);
4996 seldrain(&vi->vpi_selinfo);
4997 destroy_vpollinfo_free(vi);
4998 }
4999
5000 /*
5001 * Initialize per-vnode helper structure to hold poll-related state.
5002 */
5003 void
5004 v_addpollinfo(struct vnode *vp)
5005 {
5006 struct vpollinfo *vi;
5007
5008 if (vp->v_pollinfo != NULL)
5009 return;
5010 vi = malloc(sizeof(*vi), M_VNODEPOLL, M_WAITOK | M_ZERO);
5011 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
5012 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
5013 vfs_knlunlock, vfs_knl_assert_lock);
5014 VI_LOCK(vp);
5015 if (vp->v_pollinfo != NULL) {
5016 VI_UNLOCK(vp);
5017 destroy_vpollinfo_free(vi);
5018 return;
5019 }
5020 vp->v_pollinfo = vi;
5021 VI_UNLOCK(vp);
5022 }
5023
5024 /*
5025 * Record a process's interest in events which might happen to
5026 * a vnode. Because poll uses the historic select-style interface
5027 * internally, this routine serves as both the ``check for any
5028 * pending events'' and the ``record my interest in future events''
5029 * functions. (These are done together, while the lock is held,
5030 * to avoid race conditions.)
5031 */
5032 int
5033 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
5034 {
5035
5036 v_addpollinfo(vp);
5037 mtx_lock(&vp->v_pollinfo->vpi_lock);
5038 if (vp->v_pollinfo->vpi_revents & events) {
5039 /*
5040 * This leaves events we are not interested
5041 * in available for the other process which
5042 * which presumably had requested them
5043 * (otherwise they would never have been
5044 * recorded).
5045 */
5046 events &= vp->v_pollinfo->vpi_revents;
5047 vp->v_pollinfo->vpi_revents &= ~events;
5048
5049 mtx_unlock(&vp->v_pollinfo->vpi_lock);
5050 return (events);
5051 }
5052 vp->v_pollinfo->vpi_events |= events;
5053 selrecord(td, &vp->v_pollinfo->vpi_selinfo);
5054 mtx_unlock(&vp->v_pollinfo->vpi_lock);
5055 return (0);
5056 }
5057
5058 /*
5059 * Routine to create and manage a filesystem syncer vnode.
5060 */
5061 #define sync_close ((int (*)(struct vop_close_args *))nullop)
5062 static int sync_fsync(struct vop_fsync_args *);
5063 static int sync_inactive(struct vop_inactive_args *);
5064 static int sync_reclaim(struct vop_reclaim_args *);
5065
5066 static struct vop_vector sync_vnodeops = {
5067 .vop_bypass = VOP_EOPNOTSUPP,
5068 .vop_close = sync_close, /* close */
5069 .vop_fsync = sync_fsync, /* fsync */
5070 .vop_inactive = sync_inactive, /* inactive */
5071 .vop_need_inactive = vop_stdneed_inactive, /* need_inactive */
5072 .vop_reclaim = sync_reclaim, /* reclaim */
5073 .vop_lock1 = vop_stdlock, /* lock */
5074 .vop_unlock = vop_stdunlock, /* unlock */
5075 .vop_islocked = vop_stdislocked, /* islocked */
5076 };
5077 VFS_VOP_VECTOR_REGISTER(sync_vnodeops);
5078
5079 /*
5080 * Create a new filesystem syncer vnode for the specified mount point.
5081 */
5082 void
5083 vfs_allocate_syncvnode(struct mount *mp)
5084 {
5085 struct vnode *vp;
5086 struct bufobj *bo;
5087 static long start, incr, next;
5088 int error;
5089
5090 /* Allocate a new vnode */
5091 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
5092 if (error != 0)
5093 panic("vfs_allocate_syncvnode: getnewvnode() failed");
5094 vp->v_type = VNON;
5095 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5096 vp->v_vflag |= VV_FORCEINSMQ;
5097 error = insmntque1(vp, mp);
5098 if (error != 0)
5099 panic("vfs_allocate_syncvnode: insmntque() failed");
5100 vp->v_vflag &= ~VV_FORCEINSMQ;
5101 vn_set_state(vp, VSTATE_CONSTRUCTED);
5102 VOP_UNLOCK(vp);
5103 /*
5104 * Place the vnode onto the syncer worklist. We attempt to
5105 * scatter them about on the list so that they will go off
5106 * at evenly distributed times even if all the filesystems
5107 * are mounted at once.
5108 */
5109 next += incr;
5110 if (next == 0 || next > syncer_maxdelay) {
5111 start /= 2;
5112 incr /= 2;
5113 if (start == 0) {
5114 start = syncer_maxdelay / 2;
5115 incr = syncer_maxdelay;
5116 }
5117 next = start;
5118 }
5119 bo = &vp->v_bufobj;
5120 BO_LOCK(bo);
5121 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
5122 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
5123 mtx_lock(&sync_mtx);
5124 sync_vnode_count++;
5125 if (mp->mnt_syncer == NULL) {
5126 mp->mnt_syncer = vp;
5127 vp = NULL;
5128 }
5129 mtx_unlock(&sync_mtx);
5130 BO_UNLOCK(bo);
5131 if (vp != NULL) {
5132 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
5133 vgone(vp);
5134 vput(vp);
5135 }
5136 }
5137
5138 void
5139 vfs_deallocate_syncvnode(struct mount *mp)
5140 {
5141 struct vnode *vp;
5142
5143 mtx_lock(&sync_mtx);
5144 vp = mp->mnt_syncer;
5145 if (vp != NULL)
5146 mp->mnt_syncer = NULL;
5147 mtx_unlock(&sync_mtx);
5148 if (vp != NULL)
5149 vrele(vp);
5150 }
5151
5152 /*
5153 * Do a lazy sync of the filesystem.
5154 */
5155 static int
5156 sync_fsync(struct vop_fsync_args *ap)
5157 {
5158 struct vnode *syncvp = ap->a_vp;
5159 struct mount *mp = syncvp->v_mount;
5160 int error, save;
5161 struct bufobj *bo;
5162
5163 /*
5164 * We only need to do something if this is a lazy evaluation.
5165 */
5166 if (ap->a_waitfor != MNT_LAZY)
5167 return (0);
5168
5169 /*
5170 * Move ourselves to the back of the sync list.
5171 */
5172 bo = &syncvp->v_bufobj;
5173 BO_LOCK(bo);
5174 vn_syncer_add_to_worklist(bo, syncdelay);
5175 BO_UNLOCK(bo);
5176
5177 /*
5178 * Walk the list of vnodes pushing all that are dirty and
5179 * not already on the sync list.
5180 */
5181 if (vfs_busy(mp, MBF_NOWAIT) != 0)
5182 return (0);
5183 VOP_UNLOCK(syncvp);
5184 save = curthread_pflags_set(TDP_SYNCIO);
5185 /*
5186 * The filesystem at hand may be idle with free vnodes stored in the
5187 * batch. Return them instead of letting them stay there indefinitely.
5188 */
5189 vfs_periodic(mp, MNT_NOWAIT);
5190 error = VFS_SYNC(mp, MNT_LAZY);
5191 curthread_pflags_restore(save);
5192 vn_lock(syncvp, LK_EXCLUSIVE | LK_RETRY);
5193 vfs_unbusy(mp);
5194 return (error);
5195 }
5196
5197 /*
5198 * The syncer vnode is no referenced.
5199 */
5200 static int
5201 sync_inactive(struct vop_inactive_args *ap)
5202 {
5203
5204 vgone(ap->a_vp);
5205 return (0);
5206 }
5207
5208 /*
5209 * The syncer vnode is no longer needed and is being decommissioned.
5210 *
5211 * Modifications to the worklist must be protected by sync_mtx.
5212 */
5213 static int
5214 sync_reclaim(struct vop_reclaim_args *ap)
5215 {
5216 struct vnode *vp = ap->a_vp;
5217 struct bufobj *bo;
5218
5219 bo = &vp->v_bufobj;
5220 BO_LOCK(bo);
5221 mtx_lock(&sync_mtx);
5222 if (vp->v_mount->mnt_syncer == vp)
5223 vp->v_mount->mnt_syncer = NULL;
5224 if (bo->bo_flag & BO_ONWORKLST) {
5225 LIST_REMOVE(bo, bo_synclist);
5226 syncer_worklist_len--;
5227 sync_vnode_count--;
5228 bo->bo_flag &= ~BO_ONWORKLST;
5229 }
5230 mtx_unlock(&sync_mtx);
5231 BO_UNLOCK(bo);
5232
5233 return (0);
5234 }
5235
5236 int
5237 vn_need_pageq_flush(struct vnode *vp)
5238 {
5239 struct vm_object *obj;
5240
5241 obj = vp->v_object;
5242 return (obj != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
5243 vm_object_mightbedirty(obj));
5244 }
5245
5246 /*
5247 * Check if vnode represents a disk device
5248 */
5249 bool
5250 vn_isdisk_error(struct vnode *vp, int *errp)
5251 {
5252 int error;
5253
5254 if (vp->v_type != VCHR) {
5255 error = ENOTBLK;
5256 goto out;
5257 }
5258 error = 0;
5259 dev_lock();
5260 if (vp->v_rdev == NULL)
5261 error = ENXIO;
5262 else if (vp->v_rdev->si_devsw == NULL)
5263 error = ENXIO;
5264 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
5265 error = ENOTBLK;
5266 dev_unlock();
5267 out:
5268 *errp = error;
5269 return (error == 0);
5270 }
5271
5272 bool
5273 vn_isdisk(struct vnode *vp)
5274 {
5275 int error;
5276
5277 return (vn_isdisk_error(vp, &error));
5278 }
5279
5280 /*
5281 * VOP_FPLOOKUP_VEXEC routines are subject to special circumstances, see
5282 * the comment above cache_fplookup for details.
5283 */
5284 int
5285 vaccess_vexec_smr(mode_t file_mode, uid_t file_uid, gid_t file_gid, struct ucred *cred)
5286 {
5287 int error;
5288
5289 VFS_SMR_ASSERT_ENTERED();
5290
5291 /* Check the owner. */
5292 if (cred->cr_uid == file_uid) {
5293 if (file_mode & S_IXUSR)
5294 return (0);
5295 goto out_error;
5296 }
5297
5298 /* Otherwise, check the groups (first match) */
5299 if (groupmember(file_gid, cred)) {
5300 if (file_mode & S_IXGRP)
5301 return (0);
5302 goto out_error;
5303 }
5304
5305 /* Otherwise, check everyone else. */
5306 if (file_mode & S_IXOTH)
5307 return (0);
5308 out_error:
5309 /*
5310 * Permission check failed, but it is possible denial will get overwritten
5311 * (e.g., when root is traversing through a 700 directory owned by someone
5312 * else).
5313 *
5314 * vaccess() calls priv_check_cred which in turn can descent into MAC
5315 * modules overriding this result. It's quite unclear what semantics
5316 * are allowed for them to operate, thus for safety we don't call them
5317 * from within the SMR section. This also means if any such modules
5318 * are present, we have to let the regular lookup decide.
5319 */
5320 error = priv_check_cred_vfs_lookup_nomac(cred);
5321 switch (error) {
5322 case 0:
5323 return (0);
5324 case EAGAIN:
5325 /*
5326 * MAC modules present.
5327 */
5328 return (EAGAIN);
5329 case EPERM:
5330 return (EACCES);
5331 default:
5332 return (error);
5333 }
5334 }
5335
5336 /*
5337 * Common filesystem object access control check routine. Accepts a
5338 * vnode's type, "mode", uid and gid, requested access mode, and credentials.
5339 * Returns 0 on success, or an errno on failure.
5340 */
5341 int
5342 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
5343 accmode_t accmode, struct ucred *cred)
5344 {
5345 accmode_t dac_granted;
5346 accmode_t priv_granted;
5347
5348 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
5349 ("invalid bit in accmode"));
5350 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
5351 ("VAPPEND without VWRITE"));
5352
5353 /*
5354 * Look for a normal, non-privileged way to access the file/directory
5355 * as requested. If it exists, go with that.
5356 */
5357
5358 dac_granted = 0;
5359
5360 /* Check the owner. */
5361 if (cred->cr_uid == file_uid) {
5362 dac_granted |= VADMIN;
5363 if (file_mode & S_IXUSR)
5364 dac_granted |= VEXEC;
5365 if (file_mode & S_IRUSR)
5366 dac_granted |= VREAD;
5367 if (file_mode & S_IWUSR)
5368 dac_granted |= (VWRITE | VAPPEND);
5369
5370 if ((accmode & dac_granted) == accmode)
5371 return (0);
5372
5373 goto privcheck;
5374 }
5375
5376 /* Otherwise, check the groups (first match) */
5377 if (groupmember(file_gid, cred)) {
5378 if (file_mode & S_IXGRP)
5379 dac_granted |= VEXEC;
5380 if (file_mode & S_IRGRP)
5381 dac_granted |= VREAD;
5382 if (file_mode & S_IWGRP)
5383 dac_granted |= (VWRITE | VAPPEND);
5384
5385 if ((accmode & dac_granted) == accmode)
5386 return (0);
5387
5388 goto privcheck;
5389 }
5390
5391 /* Otherwise, check everyone else. */
5392 if (file_mode & S_IXOTH)
5393 dac_granted |= VEXEC;
5394 if (file_mode & S_IROTH)
5395 dac_granted |= VREAD;
5396 if (file_mode & S_IWOTH)
5397 dac_granted |= (VWRITE | VAPPEND);
5398 if ((accmode & dac_granted) == accmode)
5399 return (0);
5400
5401 privcheck:
5402 /*
5403 * Build a privilege mask to determine if the set of privileges
5404 * satisfies the requirements when combined with the granted mask
5405 * from above. For each privilege, if the privilege is required,
5406 * bitwise or the request type onto the priv_granted mask.
5407 */
5408 priv_granted = 0;
5409
5410 if (type == VDIR) {
5411 /*
5412 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
5413 * requests, instead of PRIV_VFS_EXEC.
5414 */
5415 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5416 !priv_check_cred(cred, PRIV_VFS_LOOKUP))
5417 priv_granted |= VEXEC;
5418 } else {
5419 /*
5420 * Ensure that at least one execute bit is on. Otherwise,
5421 * a privileged user will always succeed, and we don't want
5422 * this to happen unless the file really is executable.
5423 */
5424 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
5425 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
5426 !priv_check_cred(cred, PRIV_VFS_EXEC))
5427 priv_granted |= VEXEC;
5428 }
5429
5430 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
5431 !priv_check_cred(cred, PRIV_VFS_READ))
5432 priv_granted |= VREAD;
5433
5434 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
5435 !priv_check_cred(cred, PRIV_VFS_WRITE))
5436 priv_granted |= (VWRITE | VAPPEND);
5437
5438 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
5439 !priv_check_cred(cred, PRIV_VFS_ADMIN))
5440 priv_granted |= VADMIN;
5441
5442 if ((accmode & (priv_granted | dac_granted)) == accmode) {
5443 return (0);
5444 }
5445
5446 return ((accmode & VADMIN) ? EPERM : EACCES);
5447 }
5448
5449 /*
5450 * Credential check based on process requesting service, and per-attribute
5451 * permissions.
5452 */
5453 int
5454 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
5455 struct thread *td, accmode_t accmode)
5456 {
5457
5458 /*
5459 * Kernel-invoked always succeeds.
5460 */
5461 if (cred == NOCRED)
5462 return (0);
5463
5464 /*
5465 * Do not allow privileged processes in jail to directly manipulate
5466 * system attributes.
5467 */
5468 switch (attrnamespace) {
5469 case EXTATTR_NAMESPACE_SYSTEM:
5470 /* Potentially should be: return (EPERM); */
5471 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM));
5472 case EXTATTR_NAMESPACE_USER:
5473 return (VOP_ACCESS(vp, accmode, cred, td));
5474 default:
5475 return (EPERM);
5476 }
5477 }
5478
5479 #ifdef DEBUG_VFS_LOCKS
5480 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */
5481 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
5482 "Drop into debugger on lock violation");
5483
5484 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */
5485 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
5486 0, "Check for interlock across VOPs");
5487
5488 int vfs_badlock_print = 1; /* Print lock violations. */
5489 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
5490 0, "Print lock violations");
5491
5492 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */
5493 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
5494 0, "Print vnode details on lock violations");
5495
5496 #ifdef KDB
5497 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */
5498 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
5499 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
5500 #endif
5501
5502 static void
5503 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
5504 {
5505
5506 #ifdef KDB
5507 if (vfs_badlock_backtrace)
5508 kdb_backtrace();
5509 #endif
5510 if (vfs_badlock_vnode)
5511 vn_printf(vp, "vnode ");
5512 if (vfs_badlock_print)
5513 printf("%s: %p %s\n", str, (void *)vp, msg);
5514 if (vfs_badlock_ddb)
5515 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5516 }
5517
5518 void
5519 assert_vi_locked(struct vnode *vp, const char *str)
5520 {
5521
5522 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
5523 vfs_badlock("interlock is not locked but should be", str, vp);
5524 }
5525
5526 void
5527 assert_vi_unlocked(struct vnode *vp, const char *str)
5528 {
5529
5530 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
5531 vfs_badlock("interlock is locked but should not be", str, vp);
5532 }
5533
5534 void
5535 assert_vop_locked(struct vnode *vp, const char *str)
5536 {
5537 int locked;
5538
5539 if (KERNEL_PANICKED() || vp == NULL)
5540 return;
5541
5542 locked = VOP_ISLOCKED(vp);
5543 if (locked == 0 || locked == LK_EXCLOTHER)
5544 vfs_badlock("is not locked but should be", str, vp);
5545 }
5546
5547 void
5548 assert_vop_unlocked(struct vnode *vp, const char *str)
5549 {
5550 if (KERNEL_PANICKED() || vp == NULL)
5551 return;
5552
5553 if (VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
5554 vfs_badlock("is locked but should not be", str, vp);
5555 }
5556
5557 void
5558 assert_vop_elocked(struct vnode *vp, const char *str)
5559 {
5560 if (KERNEL_PANICKED() || vp == NULL)
5561 return;
5562
5563 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
5564 vfs_badlock("is not exclusive locked but should be", str, vp);
5565 }
5566 #endif /* DEBUG_VFS_LOCKS */
5567
5568 void
5569 vop_rename_fail(struct vop_rename_args *ap)
5570 {
5571
5572 if (ap->a_tvp != NULL)
5573 vput(ap->a_tvp);
5574 if (ap->a_tdvp == ap->a_tvp)
5575 vrele(ap->a_tdvp);
5576 else
5577 vput(ap->a_tdvp);
5578 vrele(ap->a_fdvp);
5579 vrele(ap->a_fvp);
5580 }
5581
5582 void
5583 vop_rename_pre(void *ap)
5584 {
5585 struct vop_rename_args *a = ap;
5586
5587 #ifdef DEBUG_VFS_LOCKS
5588 if (a->a_tvp)
5589 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
5590 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
5591 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
5592 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
5593
5594 /* Check the source (from). */
5595 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
5596 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
5597 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
5598 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
5599 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
5600
5601 /* Check the target. */
5602 if (a->a_tvp)
5603 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
5604 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
5605 #endif
5606 /*
5607 * It may be tempting to add vn_seqc_write_begin/end calls here and
5608 * in vop_rename_post but that's not going to work out since some
5609 * filesystems relookup vnodes mid-rename. This is probably a bug.
5610 *
5611 * For now filesystems are expected to do the relevant calls after they
5612 * decide what vnodes to operate on.
5613 */
5614 if (a->a_tdvp != a->a_fdvp)
5615 vhold(a->a_fdvp);
5616 if (a->a_tvp != a->a_fvp)
5617 vhold(a->a_fvp);
5618 vhold(a->a_tdvp);
5619 if (a->a_tvp)
5620 vhold(a->a_tvp);
5621 }
5622
5623 #ifdef DEBUG_VFS_LOCKS
5624 void
5625 vop_fplookup_vexec_debugpre(void *ap __unused)
5626 {
5627
5628 VFS_SMR_ASSERT_ENTERED();
5629 }
5630
5631 void
5632 vop_fplookup_vexec_debugpost(void *ap __unused, int rc __unused)
5633 {
5634
5635 VFS_SMR_ASSERT_ENTERED();
5636 }
5637
5638 void
5639 vop_fplookup_symlink_debugpre(void *ap __unused)
5640 {
5641
5642 VFS_SMR_ASSERT_ENTERED();
5643 }
5644
5645 void
5646 vop_fplookup_symlink_debugpost(void *ap __unused, int rc __unused)
5647 {
5648
5649 VFS_SMR_ASSERT_ENTERED();
5650 }
5651
5652 static void
5653 vop_fsync_debugprepost(struct vnode *vp, const char *name)
5654 {
5655 if (vp->v_type == VCHR)
5656 ;
5657 else if (MNT_EXTENDED_SHARED(vp->v_mount))
5658 ASSERT_VOP_LOCKED(vp, name);
5659 else
5660 ASSERT_VOP_ELOCKED(vp, name);
5661 }
5662
5663 void
5664 vop_fsync_debugpre(void *a)
5665 {
5666 struct vop_fsync_args *ap;
5667
5668 ap = a;
5669 vop_fsync_debugprepost(ap->a_vp, "fsync");
5670 }
5671
5672 void
5673 vop_fsync_debugpost(void *a, int rc __unused)
5674 {
5675 struct vop_fsync_args *ap;
5676
5677 ap = a;
5678 vop_fsync_debugprepost(ap->a_vp, "fsync");
5679 }
5680
5681 void
5682 vop_fdatasync_debugpre(void *a)
5683 {
5684 struct vop_fdatasync_args *ap;
5685
5686 ap = a;
5687 vop_fsync_debugprepost(ap->a_vp, "fsync");
5688 }
5689
5690 void
5691 vop_fdatasync_debugpost(void *a, int rc __unused)
5692 {
5693 struct vop_fdatasync_args *ap;
5694
5695 ap = a;
5696 vop_fsync_debugprepost(ap->a_vp, "fsync");
5697 }
5698
5699 void
5700 vop_strategy_debugpre(void *ap)
5701 {
5702 struct vop_strategy_args *a;
5703 struct buf *bp;
5704
5705 a = ap;
5706 bp = a->a_bp;
5707
5708 /*
5709 * Cluster ops lock their component buffers but not the IO container.
5710 */
5711 if ((bp->b_flags & B_CLUSTER) != 0)
5712 return;
5713
5714 if (!KERNEL_PANICKED() && !BUF_ISLOCKED(bp)) {
5715 if (vfs_badlock_print)
5716 printf(
5717 "VOP_STRATEGY: bp is not locked but should be\n");
5718 if (vfs_badlock_ddb)
5719 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
5720 }
5721 }
5722
5723 void
5724 vop_lock_debugpre(void *ap)
5725 {
5726 struct vop_lock1_args *a = ap;
5727
5728 if ((a->a_flags & LK_INTERLOCK) == 0)
5729 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5730 else
5731 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
5732 }
5733
5734 void
5735 vop_lock_debugpost(void *ap, int rc)
5736 {
5737 struct vop_lock1_args *a = ap;
5738
5739 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
5740 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
5741 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
5742 }
5743
5744 void
5745 vop_unlock_debugpre(void *ap)
5746 {
5747 struct vop_unlock_args *a = ap;
5748 struct vnode *vp = a->a_vp;
5749
5750 VNPASS(vn_get_state(vp) != VSTATE_UNINITIALIZED, vp);
5751 ASSERT_VOP_LOCKED(vp, "VOP_UNLOCK");
5752 }
5753
5754 void
5755 vop_need_inactive_debugpre(void *ap)
5756 {
5757 struct vop_need_inactive_args *a = ap;
5758
5759 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5760 }
5761
5762 void
5763 vop_need_inactive_debugpost(void *ap, int rc)
5764 {
5765 struct vop_need_inactive_args *a = ap;
5766
5767 ASSERT_VI_LOCKED(a->a_vp, "VOP_NEED_INACTIVE");
5768 }
5769 #endif
5770
5771 void
5772 vop_create_pre(void *ap)
5773 {
5774 struct vop_create_args *a;
5775 struct vnode *dvp;
5776
5777 a = ap;
5778 dvp = a->a_dvp;
5779 vn_seqc_write_begin(dvp);
5780 }
5781
5782 void
5783 vop_create_post(void *ap, int rc)
5784 {
5785 struct vop_create_args *a;
5786 struct vnode *dvp;
5787
5788 a = ap;
5789 dvp = a->a_dvp;
5790 vn_seqc_write_end(dvp);
5791 if (!rc)
5792 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5793 }
5794
5795 void
5796 vop_whiteout_pre(void *ap)
5797 {
5798 struct vop_whiteout_args *a;
5799 struct vnode *dvp;
5800
5801 a = ap;
5802 dvp = a->a_dvp;
5803 vn_seqc_write_begin(dvp);
5804 }
5805
5806 void
5807 vop_whiteout_post(void *ap, int rc)
5808 {
5809 struct vop_whiteout_args *a;
5810 struct vnode *dvp;
5811
5812 a = ap;
5813 dvp = a->a_dvp;
5814 vn_seqc_write_end(dvp);
5815 }
5816
5817 void
5818 vop_deleteextattr_pre(void *ap)
5819 {
5820 struct vop_deleteextattr_args *a;
5821 struct vnode *vp;
5822
5823 a = ap;
5824 vp = a->a_vp;
5825 vn_seqc_write_begin(vp);
5826 }
5827
5828 void
5829 vop_deleteextattr_post(void *ap, int rc)
5830 {
5831 struct vop_deleteextattr_args *a;
5832 struct vnode *vp;
5833
5834 a = ap;
5835 vp = a->a_vp;
5836 vn_seqc_write_end(vp);
5837 if (!rc)
5838 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
5839 }
5840
5841 void
5842 vop_link_pre(void *ap)
5843 {
5844 struct vop_link_args *a;
5845 struct vnode *vp, *tdvp;
5846
5847 a = ap;
5848 vp = a->a_vp;
5849 tdvp = a->a_tdvp;
5850 vn_seqc_write_begin(vp);
5851 vn_seqc_write_begin(tdvp);
5852 }
5853
5854 void
5855 vop_link_post(void *ap, int rc)
5856 {
5857 struct vop_link_args *a;
5858 struct vnode *vp, *tdvp;
5859
5860 a = ap;
5861 vp = a->a_vp;
5862 tdvp = a->a_tdvp;
5863 vn_seqc_write_end(vp);
5864 vn_seqc_write_end(tdvp);
5865 if (!rc) {
5866 VFS_KNOTE_LOCKED(vp, NOTE_LINK);
5867 VFS_KNOTE_LOCKED(tdvp, NOTE_WRITE);
5868 }
5869 }
5870
5871 void
5872 vop_mkdir_pre(void *ap)
5873 {
5874 struct vop_mkdir_args *a;
5875 struct vnode *dvp;
5876
5877 a = ap;
5878 dvp = a->a_dvp;
5879 vn_seqc_write_begin(dvp);
5880 }
5881
5882 void
5883 vop_mkdir_post(void *ap, int rc)
5884 {
5885 struct vop_mkdir_args *a;
5886 struct vnode *dvp;
5887
5888 a = ap;
5889 dvp = a->a_dvp;
5890 vn_seqc_write_end(dvp);
5891 if (!rc)
5892 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
5893 }
5894
5895 #ifdef DEBUG_VFS_LOCKS
5896 void
5897 vop_mkdir_debugpost(void *ap, int rc)
5898 {
5899 struct vop_mkdir_args *a;
5900
5901 a = ap;
5902 if (!rc)
5903 cache_validate(a->a_dvp, *a->a_vpp, a->a_cnp);
5904 }
5905 #endif
5906
5907 void
5908 vop_mknod_pre(void *ap)
5909 {
5910 struct vop_mknod_args *a;
5911 struct vnode *dvp;
5912
5913 a = ap;
5914 dvp = a->a_dvp;
5915 vn_seqc_write_begin(dvp);
5916 }
5917
5918 void
5919 vop_mknod_post(void *ap, int rc)
5920 {
5921 struct vop_mknod_args *a;
5922 struct vnode *dvp;
5923
5924 a = ap;
5925 dvp = a->a_dvp;
5926 vn_seqc_write_end(dvp);
5927 if (!rc)
5928 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5929 }
5930
5931 void
5932 vop_reclaim_post(void *ap, int rc)
5933 {
5934 struct vop_reclaim_args *a;
5935 struct vnode *vp;
5936
5937 a = ap;
5938 vp = a->a_vp;
5939 ASSERT_VOP_IN_SEQC(vp);
5940 if (!rc)
5941 VFS_KNOTE_LOCKED(vp, NOTE_REVOKE);
5942 }
5943
5944 void
5945 vop_remove_pre(void *ap)
5946 {
5947 struct vop_remove_args *a;
5948 struct vnode *dvp, *vp;
5949
5950 a = ap;
5951 dvp = a->a_dvp;
5952 vp = a->a_vp;
5953 vn_seqc_write_begin(dvp);
5954 vn_seqc_write_begin(vp);
5955 }
5956
5957 void
5958 vop_remove_post(void *ap, int rc)
5959 {
5960 struct vop_remove_args *a;
5961 struct vnode *dvp, *vp;
5962
5963 a = ap;
5964 dvp = a->a_dvp;
5965 vp = a->a_vp;
5966 vn_seqc_write_end(dvp);
5967 vn_seqc_write_end(vp);
5968 if (!rc) {
5969 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
5970 VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
5971 }
5972 }
5973
5974 void
5975 vop_rename_post(void *ap, int rc)
5976 {
5977 struct vop_rename_args *a = ap;
5978 long hint;
5979
5980 if (!rc) {
5981 hint = NOTE_WRITE;
5982 if (a->a_fdvp == a->a_tdvp) {
5983 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
5984 hint |= NOTE_LINK;
5985 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5986 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5987 } else {
5988 hint |= NOTE_EXTEND;
5989 if (a->a_fvp->v_type == VDIR)
5990 hint |= NOTE_LINK;
5991 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
5992
5993 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
5994 a->a_tvp->v_type == VDIR)
5995 hint &= ~NOTE_LINK;
5996 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
5997 }
5998
5999 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
6000 if (a->a_tvp)
6001 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
6002 }
6003 if (a->a_tdvp != a->a_fdvp)
6004 vdrop(a->a_fdvp);
6005 if (a->a_tvp != a->a_fvp)
6006 vdrop(a->a_fvp);
6007 vdrop(a->a_tdvp);
6008 if (a->a_tvp)
6009 vdrop(a->a_tvp);
6010 }
6011
6012 void
6013 vop_rmdir_pre(void *ap)
6014 {
6015 struct vop_rmdir_args *a;
6016 struct vnode *dvp, *vp;
6017
6018 a = ap;
6019 dvp = a->a_dvp;
6020 vp = a->a_vp;
6021 vn_seqc_write_begin(dvp);
6022 vn_seqc_write_begin(vp);
6023 }
6024
6025 void
6026 vop_rmdir_post(void *ap, int rc)
6027 {
6028 struct vop_rmdir_args *a;
6029 struct vnode *dvp, *vp;
6030
6031 a = ap;
6032 dvp = a->a_dvp;
6033 vp = a->a_vp;
6034 vn_seqc_write_end(dvp);
6035 vn_seqc_write_end(vp);
6036 if (!rc) {
6037 vp->v_vflag |= VV_UNLINKED;
6038 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE | NOTE_LINK);
6039 VFS_KNOTE_LOCKED(vp, NOTE_DELETE);
6040 }
6041 }
6042
6043 void
6044 vop_setattr_pre(void *ap)
6045 {
6046 struct vop_setattr_args *a;
6047 struct vnode *vp;
6048
6049 a = ap;
6050 vp = a->a_vp;
6051 vn_seqc_write_begin(vp);
6052 }
6053
6054 void
6055 vop_setattr_post(void *ap, int rc)
6056 {
6057 struct vop_setattr_args *a;
6058 struct vnode *vp;
6059
6060 a = ap;
6061 vp = a->a_vp;
6062 vn_seqc_write_end(vp);
6063 if (!rc)
6064 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6065 }
6066
6067 void
6068 vop_setacl_pre(void *ap)
6069 {
6070 struct vop_setacl_args *a;
6071 struct vnode *vp;
6072
6073 a = ap;
6074 vp = a->a_vp;
6075 vn_seqc_write_begin(vp);
6076 }
6077
6078 void
6079 vop_setacl_post(void *ap, int rc __unused)
6080 {
6081 struct vop_setacl_args *a;
6082 struct vnode *vp;
6083
6084 a = ap;
6085 vp = a->a_vp;
6086 vn_seqc_write_end(vp);
6087 }
6088
6089 void
6090 vop_setextattr_pre(void *ap)
6091 {
6092 struct vop_setextattr_args *a;
6093 struct vnode *vp;
6094
6095 a = ap;
6096 vp = a->a_vp;
6097 vn_seqc_write_begin(vp);
6098 }
6099
6100 void
6101 vop_setextattr_post(void *ap, int rc)
6102 {
6103 struct vop_setextattr_args *a;
6104 struct vnode *vp;
6105
6106 a = ap;
6107 vp = a->a_vp;
6108 vn_seqc_write_end(vp);
6109 if (!rc)
6110 VFS_KNOTE_LOCKED(vp, NOTE_ATTRIB);
6111 }
6112
6113 void
6114 vop_symlink_pre(void *ap)
6115 {
6116 struct vop_symlink_args *a;
6117 struct vnode *dvp;
6118
6119 a = ap;
6120 dvp = a->a_dvp;
6121 vn_seqc_write_begin(dvp);
6122 }
6123
6124 void
6125 vop_symlink_post(void *ap, int rc)
6126 {
6127 struct vop_symlink_args *a;
6128 struct vnode *dvp;
6129
6130 a = ap;
6131 dvp = a->a_dvp;
6132 vn_seqc_write_end(dvp);
6133 if (!rc)
6134 VFS_KNOTE_LOCKED(dvp, NOTE_WRITE);
6135 }
6136
6137 void
6138 vop_open_post(void *ap, int rc)
6139 {
6140 struct vop_open_args *a = ap;
6141
6142 if (!rc)
6143 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
6144 }
6145
6146 void
6147 vop_close_post(void *ap, int rc)
6148 {
6149 struct vop_close_args *a = ap;
6150
6151 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
6152 !VN_IS_DOOMED(a->a_vp))) {
6153 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
6154 NOTE_CLOSE_WRITE : NOTE_CLOSE);
6155 }
6156 }
6157
6158 void
6159 vop_read_post(void *ap, int rc)
6160 {
6161 struct vop_read_args *a = ap;
6162
6163 if (!rc)
6164 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6165 }
6166
6167 void
6168 vop_read_pgcache_post(void *ap, int rc)
6169 {
6170 struct vop_read_pgcache_args *a = ap;
6171
6172 if (!rc)
6173 VFS_KNOTE_UNLOCKED(a->a_vp, NOTE_READ);
6174 }
6175
6176 void
6177 vop_readdir_post(void *ap, int rc)
6178 {
6179 struct vop_readdir_args *a = ap;
6180
6181 if (!rc)
6182 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
6183 }
6184
6185 static struct knlist fs_knlist;
6186
6187 static void
6188 vfs_event_init(void *arg)
6189 {
6190 knlist_init_mtx(&fs_knlist, NULL);
6191 }
6192 /* XXX - correct order? */
6193 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
6194
6195 void
6196 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
6197 {
6198
6199 KNOTE_UNLOCKED(&fs_knlist, event);
6200 }
6201
6202 static int filt_fsattach(struct knote *kn);
6203 static void filt_fsdetach(struct knote *kn);
6204 static int filt_fsevent(struct knote *kn, long hint);
6205
6206 struct filterops fs_filtops = {
6207 .f_isfd = 0,
6208 .f_attach = filt_fsattach,
6209 .f_detach = filt_fsdetach,
6210 .f_event = filt_fsevent
6211 };
6212
6213 static int
6214 filt_fsattach(struct knote *kn)
6215 {
6216
6217 kn->kn_flags |= EV_CLEAR;
6218 knlist_add(&fs_knlist, kn, 0);
6219 return (0);
6220 }
6221
6222 static void
6223 filt_fsdetach(struct knote *kn)
6224 {
6225
6226 knlist_remove(&fs_knlist, kn, 0);
6227 }
6228
6229 static int
6230 filt_fsevent(struct knote *kn, long hint)
6231 {
6232
6233 kn->kn_fflags |= kn->kn_sfflags & hint;
6234
6235 return (kn->kn_fflags != 0);
6236 }
6237
6238 static int
6239 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
6240 {
6241 struct vfsidctl vc;
6242 int error;
6243 struct mount *mp;
6244
6245 error = SYSCTL_IN(req, &vc, sizeof(vc));
6246 if (error)
6247 return (error);
6248 if (vc.vc_vers != VFS_CTL_VERS1)
6249 return (EINVAL);
6250 mp = vfs_getvfs(&vc.vc_fsid);
6251 if (mp == NULL)
6252 return (ENOENT);
6253 /* ensure that a specific sysctl goes to the right filesystem. */
6254 if (strcmp(vc.vc_fstypename, "*") != 0 &&
6255 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
6256 vfs_rel(mp);
6257 return (EINVAL);
6258 }
6259 VCTLTOREQ(&vc, req);
6260 error = VFS_SYSCTL(mp, vc.vc_op, req);
6261 vfs_rel(mp);
6262 return (error);
6263 }
6264
6265 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_MPSAFE | CTLFLAG_WR,
6266 NULL, 0, sysctl_vfs_ctl, "",
6267 "Sysctl by fsid");
6268
6269 /*
6270 * Function to initialize a va_filerev field sensibly.
6271 * XXX: Wouldn't a random number make a lot more sense ??
6272 */
6273 u_quad_t
6274 init_va_filerev(void)
6275 {
6276 struct bintime bt;
6277
6278 getbinuptime(&bt);
6279 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
6280 }
6281
6282 static int filt_vfsread(struct knote *kn, long hint);
6283 static int filt_vfswrite(struct knote *kn, long hint);
6284 static int filt_vfsvnode(struct knote *kn, long hint);
6285 static void filt_vfsdetach(struct knote *kn);
6286 static struct filterops vfsread_filtops = {
6287 .f_isfd = 1,
6288 .f_detach = filt_vfsdetach,
6289 .f_event = filt_vfsread
6290 };
6291 static struct filterops vfswrite_filtops = {
6292 .f_isfd = 1,
6293 .f_detach = filt_vfsdetach,
6294 .f_event = filt_vfswrite
6295 };
6296 static struct filterops vfsvnode_filtops = {
6297 .f_isfd = 1,
6298 .f_detach = filt_vfsdetach,
6299 .f_event = filt_vfsvnode
6300 };
6301
6302 static void
6303 vfs_knllock(void *arg)
6304 {
6305 struct vnode *vp = arg;
6306
6307 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
6308 }
6309
6310 static void
6311 vfs_knlunlock(void *arg)
6312 {
6313 struct vnode *vp = arg;
6314
6315 VOP_UNLOCK(vp);
6316 }
6317
6318 static void
6319 vfs_knl_assert_lock(void *arg, int what)
6320 {
6321 #ifdef DEBUG_VFS_LOCKS
6322 struct vnode *vp = arg;
6323
6324 if (what == LA_LOCKED)
6325 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
6326 else
6327 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
6328 #endif
6329 }
6330
6331 int
6332 vfs_kqfilter(struct vop_kqfilter_args *ap)
6333 {
6334 struct vnode *vp = ap->a_vp;
6335 struct knote *kn = ap->a_kn;
6336 struct knlist *knl;
6337
6338 KASSERT(vp->v_type != VFIFO || (kn->kn_filter != EVFILT_READ &&
6339 kn->kn_filter != EVFILT_WRITE),
6340 ("READ/WRITE filter on a FIFO leaked through"));
6341 switch (kn->kn_filter) {
6342 case EVFILT_READ:
6343 kn->kn_fop = &vfsread_filtops;
6344 break;
6345 case EVFILT_WRITE:
6346 kn->kn_fop = &vfswrite_filtops;
6347 break;
6348 case EVFILT_VNODE:
6349 kn->kn_fop = &vfsvnode_filtops;
6350 break;
6351 default:
6352 return (EINVAL);
6353 }
6354
6355 kn->kn_hook = (caddr_t)vp;
6356
6357 v_addpollinfo(vp);
6358 if (vp->v_pollinfo == NULL)
6359 return (ENOMEM);
6360 knl = &vp->v_pollinfo->vpi_selinfo.si_note;
6361 vhold(vp);
6362 knlist_add(knl, kn, 0);
6363
6364 return (0);
6365 }
6366
6367 /*
6368 * Detach knote from vnode
6369 */
6370 static void
6371 filt_vfsdetach(struct knote *kn)
6372 {
6373 struct vnode *vp = (struct vnode *)kn->kn_hook;
6374
6375 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
6376 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
6377 vdrop(vp);
6378 }
6379
6380 /*ARGSUSED*/
6381 static int
6382 filt_vfsread(struct knote *kn, long hint)
6383 {
6384 struct vnode *vp = (struct vnode *)kn->kn_hook;
6385 off_t size;
6386 int res;
6387
6388 /*
6389 * filesystem is gone, so set the EOF flag and schedule
6390 * the knote for deletion.
6391 */
6392 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6393 VI_LOCK(vp);
6394 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6395 VI_UNLOCK(vp);
6396 return (1);
6397 }
6398
6399 if (vn_getsize_locked(vp, &size, curthread->td_ucred) != 0)
6400 return (0);
6401
6402 VI_LOCK(vp);
6403 kn->kn_data = size - kn->kn_fp->f_offset;
6404 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
6405 VI_UNLOCK(vp);
6406 return (res);
6407 }
6408
6409 /*ARGSUSED*/
6410 static int
6411 filt_vfswrite(struct knote *kn, long hint)
6412 {
6413 struct vnode *vp = (struct vnode *)kn->kn_hook;
6414
6415 VI_LOCK(vp);
6416
6417 /*
6418 * filesystem is gone, so set the EOF flag and schedule
6419 * the knote for deletion.
6420 */
6421 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
6422 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
6423
6424 kn->kn_data = 0;
6425 VI_UNLOCK(vp);
6426 return (1);
6427 }
6428
6429 static int
6430 filt_vfsvnode(struct knote *kn, long hint)
6431 {
6432 struct vnode *vp = (struct vnode *)kn->kn_hook;
6433 int res;
6434
6435 VI_LOCK(vp);
6436 if (kn->kn_sfflags & hint)
6437 kn->kn_fflags |= hint;
6438 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
6439 kn->kn_flags |= EV_EOF;
6440 VI_UNLOCK(vp);
6441 return (1);
6442 }
6443 res = (kn->kn_fflags != 0);
6444 VI_UNLOCK(vp);
6445 return (res);
6446 }
6447
6448 /*
6449 * Returns whether the directory is empty or not.
6450 * If it is empty, the return value is 0; otherwise
6451 * the return value is an error value (which may
6452 * be ENOTEMPTY).
6453 */
6454 int
6455 vfs_emptydir(struct vnode *vp)
6456 {
6457 struct uio uio;
6458 struct iovec iov;
6459 struct dirent *dirent, *dp, *endp;
6460 int error, eof;
6461
6462 error = 0;
6463 eof = 0;
6464
6465 ASSERT_VOP_LOCKED(vp, "vfs_emptydir");
6466 VNASSERT(vp->v_type == VDIR, vp, ("vp is not a directory"));
6467
6468 dirent = malloc(sizeof(struct dirent), M_TEMP, M_WAITOK);
6469 iov.iov_base = dirent;
6470 iov.iov_len = sizeof(struct dirent);
6471
6472 uio.uio_iov = &iov;
6473 uio.uio_iovcnt = 1;
6474 uio.uio_offset = 0;
6475 uio.uio_resid = sizeof(struct dirent);
6476 uio.uio_segflg = UIO_SYSSPACE;
6477 uio.uio_rw = UIO_READ;
6478 uio.uio_td = curthread;
6479
6480 while (eof == 0 && error == 0) {
6481 error = VOP_READDIR(vp, &uio, curthread->td_ucred, &eof,
6482 NULL, NULL);
6483 if (error != 0)
6484 break;
6485 endp = (void *)((uint8_t *)dirent +
6486 sizeof(struct dirent) - uio.uio_resid);
6487 for (dp = dirent; dp < endp;
6488 dp = (void *)((uint8_t *)dp + GENERIC_DIRSIZ(dp))) {
6489 if (dp->d_type == DT_WHT)
6490 continue;
6491 if (dp->d_namlen == 0)
6492 continue;
6493 if (dp->d_type != DT_DIR &&
6494 dp->d_type != DT_UNKNOWN) {
6495 error = ENOTEMPTY;
6496 break;
6497 }
6498 if (dp->d_namlen > 2) {
6499 error = ENOTEMPTY;
6500 break;
6501 }
6502 if (dp->d_namlen == 1 &&
6503 dp->d_name[0] != '.') {
6504 error = ENOTEMPTY;
6505 break;
6506 }
6507 if (dp->d_namlen == 2 &&
6508 dp->d_name[1] != '.') {
6509 error = ENOTEMPTY;
6510 break;
6511 }
6512 uio.uio_resid = sizeof(struct dirent);
6513 }
6514 }
6515 free(dirent, M_TEMP);
6516 return (error);
6517 }
6518
6519 int
6520 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
6521 {
6522 int error;
6523
6524 if (dp->d_reclen > ap->a_uio->uio_resid)
6525 return (ENAMETOOLONG);
6526 error = uiomove(dp, dp->d_reclen, ap->a_uio);
6527 if (error) {
6528 if (ap->a_ncookies != NULL) {
6529 if (ap->a_cookies != NULL)
6530 free(ap->a_cookies, M_TEMP);
6531 ap->a_cookies = NULL;
6532 *ap->a_ncookies = 0;
6533 }
6534 return (error);
6535 }
6536 if (ap->a_ncookies == NULL)
6537 return (0);
6538
6539 KASSERT(ap->a_cookies,
6540 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
6541
6542 *ap->a_cookies = realloc(*ap->a_cookies,
6543 (*ap->a_ncookies + 1) * sizeof(uint64_t), M_TEMP, M_WAITOK | M_ZERO);
6544 (*ap->a_cookies)[*ap->a_ncookies] = off;
6545 *ap->a_ncookies += 1;
6546 return (0);
6547 }
6548
6549 /*
6550 * The purpose of this routine is to remove granularity from accmode_t,
6551 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
6552 * VADMIN and VAPPEND.
6553 *
6554 * If it returns 0, the caller is supposed to continue with the usual
6555 * access checks using 'accmode' as modified by this routine. If it
6556 * returns nonzero value, the caller is supposed to return that value
6557 * as errno.
6558 *
6559 * Note that after this routine runs, accmode may be zero.
6560 */
6561 int
6562 vfs_unixify_accmode(accmode_t *accmode)
6563 {
6564 /*
6565 * There is no way to specify explicit "deny" rule using
6566 * file mode or POSIX.1e ACLs.
6567 */
6568 if (*accmode & VEXPLICIT_DENY) {
6569 *accmode = 0;
6570 return (0);
6571 }
6572
6573 /*
6574 * None of these can be translated into usual access bits.
6575 * Also, the common case for NFSv4 ACLs is to not contain
6576 * either of these bits. Caller should check for VWRITE
6577 * on the containing directory instead.
6578 */
6579 if (*accmode & (VDELETE_CHILD | VDELETE))
6580 return (EPERM);
6581
6582 if (*accmode & VADMIN_PERMS) {
6583 *accmode &= ~VADMIN_PERMS;
6584 *accmode |= VADMIN;
6585 }
6586
6587 /*
6588 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
6589 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
6590 */
6591 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
6592
6593 return (0);
6594 }
6595
6596 /*
6597 * Clear out a doomed vnode (if any) and replace it with a new one as long
6598 * as the fs is not being unmounted. Return the root vnode to the caller.
6599 */
6600 static int __noinline
6601 vfs_cache_root_fallback(struct mount *mp, int flags, struct vnode **vpp)
6602 {
6603 struct vnode *vp;
6604 int error;
6605
6606 restart:
6607 if (mp->mnt_rootvnode != NULL) {
6608 MNT_ILOCK(mp);
6609 vp = mp->mnt_rootvnode;
6610 if (vp != NULL) {
6611 if (!VN_IS_DOOMED(vp)) {
6612 vrefact(vp);
6613 MNT_IUNLOCK(mp);
6614 error = vn_lock(vp, flags);
6615 if (error == 0) {
6616 *vpp = vp;
6617 return (0);
6618 }
6619 vrele(vp);
6620 goto restart;
6621 }
6622 /*
6623 * Clear the old one.
6624 */
6625 mp->mnt_rootvnode = NULL;
6626 }
6627 MNT_IUNLOCK(mp);
6628 if (vp != NULL) {
6629 vfs_op_barrier_wait(mp);
6630 vrele(vp);
6631 }
6632 }
6633 error = VFS_CACHEDROOT(mp, flags, vpp);
6634 if (error != 0)
6635 return (error);
6636 if (mp->mnt_vfs_ops == 0) {
6637 MNT_ILOCK(mp);
6638 if (mp->mnt_vfs_ops != 0) {
6639 MNT_IUNLOCK(mp);
6640 return (0);
6641 }
6642 if (mp->mnt_rootvnode == NULL) {
6643 vrefact(*vpp);
6644 mp->mnt_rootvnode = *vpp;
6645 } else {
6646 if (mp->mnt_rootvnode != *vpp) {
6647 if (!VN_IS_DOOMED(mp->mnt_rootvnode)) {
6648 panic("%s: mismatch between vnode returned "
6649 " by VFS_CACHEDROOT and the one cached "
6650 " (%p != %p)",
6651 __func__, *vpp, mp->mnt_rootvnode);
6652 }
6653 }
6654 }
6655 MNT_IUNLOCK(mp);
6656 }
6657 return (0);
6658 }
6659
6660 int
6661 vfs_cache_root(struct mount *mp, int flags, struct vnode **vpp)
6662 {
6663 struct mount_pcpu *mpcpu;
6664 struct vnode *vp;
6665 int error;
6666
6667 if (!vfs_op_thread_enter(mp, mpcpu))
6668 return (vfs_cache_root_fallback(mp, flags, vpp));
6669 vp = atomic_load_ptr(&mp->mnt_rootvnode);
6670 if (vp == NULL || VN_IS_DOOMED(vp)) {
6671 vfs_op_thread_exit(mp, mpcpu);
6672 return (vfs_cache_root_fallback(mp, flags, vpp));
6673 }
6674 vrefact(vp);
6675 vfs_op_thread_exit(mp, mpcpu);
6676 error = vn_lock(vp, flags);
6677 if (error != 0) {
6678 vrele(vp);
6679 return (vfs_cache_root_fallback(mp, flags, vpp));
6680 }
6681 *vpp = vp;
6682 return (0);
6683 }
6684
6685 struct vnode *
6686 vfs_cache_root_clear(struct mount *mp)
6687 {
6688 struct vnode *vp;
6689
6690 /*
6691 * ops > 0 guarantees there is nobody who can see this vnode
6692 */
6693 MPASS(mp->mnt_vfs_ops > 0);
6694 vp = mp->mnt_rootvnode;
6695 if (vp != NULL)
6696 vn_seqc_write_begin(vp);
6697 mp->mnt_rootvnode = NULL;
6698 return (vp);
6699 }
6700
6701 void
6702 vfs_cache_root_set(struct mount *mp, struct vnode *vp)
6703 {
6704
6705 MPASS(mp->mnt_vfs_ops > 0);
6706 vrefact(vp);
6707 mp->mnt_rootvnode = vp;
6708 }
6709
6710 /*
6711 * These are helper functions for filesystems to traverse all
6712 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
6713 *
6714 * This interface replaces MNT_VNODE_FOREACH.
6715 */
6716
6717 struct vnode *
6718 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
6719 {
6720 struct vnode *vp;
6721
6722 if (should_yield())
6723 kern_yield(PRI_USER);
6724 MNT_ILOCK(mp);
6725 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6726 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
6727 vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
6728 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6729 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6730 continue;
6731 VI_LOCK(vp);
6732 if (VN_IS_DOOMED(vp)) {
6733 VI_UNLOCK(vp);
6734 continue;
6735 }
6736 break;
6737 }
6738 if (vp == NULL) {
6739 __mnt_vnode_markerfree_all(mvp, mp);
6740 /* MNT_IUNLOCK(mp); -- done in above function */
6741 mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
6742 return (NULL);
6743 }
6744 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6745 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6746 MNT_IUNLOCK(mp);
6747 return (vp);
6748 }
6749
6750 struct vnode *
6751 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
6752 {
6753 struct vnode *vp;
6754
6755 *mvp = vn_alloc_marker(mp);
6756 MNT_ILOCK(mp);
6757 MNT_REF(mp);
6758
6759 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
6760 /* Allow a racy peek at VIRF_DOOMED to save a lock acquisition. */
6761 if (vp->v_type == VMARKER || VN_IS_DOOMED(vp))
6762 continue;
6763 VI_LOCK(vp);
6764 if (VN_IS_DOOMED(vp)) {
6765 VI_UNLOCK(vp);
6766 continue;
6767 }
6768 break;
6769 }
6770 if (vp == NULL) {
6771 MNT_REL(mp);
6772 MNT_IUNLOCK(mp);
6773 vn_free_marker(*mvp);
6774 *mvp = NULL;
6775 return (NULL);
6776 }
6777 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
6778 MNT_IUNLOCK(mp);
6779 return (vp);
6780 }
6781
6782 void
6783 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
6784 {
6785
6786 if (*mvp == NULL) {
6787 MNT_IUNLOCK(mp);
6788 return;
6789 }
6790
6791 mtx_assert(MNT_MTX(mp), MA_OWNED);
6792
6793 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6794 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
6795 MNT_REL(mp);
6796 MNT_IUNLOCK(mp);
6797 vn_free_marker(*mvp);
6798 *mvp = NULL;
6799 }
6800
6801 /*
6802 * These are helper functions for filesystems to traverse their
6803 * lazy vnodes. See MNT_VNODE_FOREACH_LAZY() in sys/mount.h
6804 */
6805 static void
6806 mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6807 {
6808
6809 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6810
6811 MNT_ILOCK(mp);
6812 MNT_REL(mp);
6813 MNT_IUNLOCK(mp);
6814 vn_free_marker(*mvp);
6815 *mvp = NULL;
6816 }
6817
6818 /*
6819 * Relock the mp mount vnode list lock with the vp vnode interlock in the
6820 * conventional lock order during mnt_vnode_next_lazy iteration.
6821 *
6822 * On entry, the mount vnode list lock is held and the vnode interlock is not.
6823 * The list lock is dropped and reacquired. On success, both locks are held.
6824 * On failure, the mount vnode list lock is held but the vnode interlock is
6825 * not, and the procedure may have yielded.
6826 */
6827 static bool
6828 mnt_vnode_next_lazy_relock(struct vnode *mvp, struct mount *mp,
6829 struct vnode *vp)
6830 {
6831
6832 VNASSERT(mvp->v_mount == mp && mvp->v_type == VMARKER &&
6833 TAILQ_NEXT(mvp, v_lazylist) != NULL, mvp,
6834 ("%s: bad marker", __func__));
6835 VNASSERT(vp->v_mount == mp && vp->v_type != VMARKER, vp,
6836 ("%s: inappropriate vnode", __func__));
6837 ASSERT_VI_UNLOCKED(vp, __func__);
6838 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6839
6840 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, mvp, v_lazylist);
6841 TAILQ_INSERT_BEFORE(vp, mvp, v_lazylist);
6842
6843 /*
6844 * Note we may be racing against vdrop which transitioned the hold
6845 * count to 0 and now waits for the ->mnt_listmtx lock. This is fine,
6846 * if we are the only user after we get the interlock we will just
6847 * vdrop.
6848 */
6849 vhold(vp);
6850 mtx_unlock(&mp->mnt_listmtx);
6851 VI_LOCK(vp);
6852 if (VN_IS_DOOMED(vp)) {
6853 VNPASS((vp->v_mflag & VMP_LAZYLIST) == 0, vp);
6854 goto out_lost;
6855 }
6856 VNPASS(vp->v_mflag & VMP_LAZYLIST, vp);
6857 /*
6858 * There is nothing to do if we are the last user.
6859 */
6860 if (!refcount_release_if_not_last(&vp->v_holdcnt))
6861 goto out_lost;
6862 mtx_lock(&mp->mnt_listmtx);
6863 return (true);
6864 out_lost:
6865 vdropl(vp);
6866 maybe_yield();
6867 mtx_lock(&mp->mnt_listmtx);
6868 return (false);
6869 }
6870
6871 static struct vnode *
6872 mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6873 void *cbarg)
6874 {
6875 struct vnode *vp;
6876
6877 mtx_assert(&mp->mnt_listmtx, MA_OWNED);
6878 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
6879 restart:
6880 vp = TAILQ_NEXT(*mvp, v_lazylist);
6881 while (vp != NULL) {
6882 if (vp->v_type == VMARKER) {
6883 vp = TAILQ_NEXT(vp, v_lazylist);
6884 continue;
6885 }
6886 /*
6887 * See if we want to process the vnode. Note we may encounter a
6888 * long string of vnodes we don't care about and hog the list
6889 * as a result. Check for it and requeue the marker.
6890 */
6891 VNPASS(!VN_IS_DOOMED(vp), vp);
6892 if (!cb(vp, cbarg)) {
6893 if (!should_yield()) {
6894 vp = TAILQ_NEXT(vp, v_lazylist);
6895 continue;
6896 }
6897 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp,
6898 v_lazylist);
6899 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp,
6900 v_lazylist);
6901 mtx_unlock(&mp->mnt_listmtx);
6902 kern_yield(PRI_USER);
6903 mtx_lock(&mp->mnt_listmtx);
6904 goto restart;
6905 }
6906 /*
6907 * Try-lock because this is the wrong lock order.
6908 */
6909 if (!VI_TRYLOCK(vp) &&
6910 !mnt_vnode_next_lazy_relock(*mvp, mp, vp))
6911 goto restart;
6912 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
6913 KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
6914 ("alien vnode on the lazy list %p %p", vp, mp));
6915 VNPASS(vp->v_mount == mp, vp);
6916 VNPASS(!VN_IS_DOOMED(vp), vp);
6917 break;
6918 }
6919 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6920
6921 /* Check if we are done */
6922 if (vp == NULL) {
6923 mtx_unlock(&mp->mnt_listmtx);
6924 mnt_vnode_markerfree_lazy(mvp, mp);
6925 return (NULL);
6926 }
6927 TAILQ_INSERT_AFTER(&mp->mnt_lazyvnodelist, vp, *mvp, v_lazylist);
6928 mtx_unlock(&mp->mnt_listmtx);
6929 ASSERT_VI_LOCKED(vp, "lazy iter");
6930 return (vp);
6931 }
6932
6933 struct vnode *
6934 __mnt_vnode_next_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6935 void *cbarg)
6936 {
6937
6938 if (should_yield())
6939 kern_yield(PRI_USER);
6940 mtx_lock(&mp->mnt_listmtx);
6941 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6942 }
6943
6944 struct vnode *
6945 __mnt_vnode_first_lazy(struct vnode **mvp, struct mount *mp, mnt_lazy_cb_t *cb,
6946 void *cbarg)
6947 {
6948 struct vnode *vp;
6949
6950 if (TAILQ_EMPTY(&mp->mnt_lazyvnodelist))
6951 return (NULL);
6952
6953 *mvp = vn_alloc_marker(mp);
6954 MNT_ILOCK(mp);
6955 MNT_REF(mp);
6956 MNT_IUNLOCK(mp);
6957
6958 mtx_lock(&mp->mnt_listmtx);
6959 vp = TAILQ_FIRST(&mp->mnt_lazyvnodelist);
6960 if (vp == NULL) {
6961 mtx_unlock(&mp->mnt_listmtx);
6962 mnt_vnode_markerfree_lazy(mvp, mp);
6963 return (NULL);
6964 }
6965 TAILQ_INSERT_BEFORE(vp, *mvp, v_lazylist);
6966 return (mnt_vnode_next_lazy(mvp, mp, cb, cbarg));
6967 }
6968
6969 void
6970 __mnt_vnode_markerfree_lazy(struct vnode **mvp, struct mount *mp)
6971 {
6972
6973 if (*mvp == NULL)
6974 return;
6975
6976 mtx_lock(&mp->mnt_listmtx);
6977 TAILQ_REMOVE(&mp->mnt_lazyvnodelist, *mvp, v_lazylist);
6978 mtx_unlock(&mp->mnt_listmtx);
6979 mnt_vnode_markerfree_lazy(mvp, mp);
6980 }
6981
6982 int
6983 vn_dir_check_exec(struct vnode *vp, struct componentname *cnp)
6984 {
6985
6986 if ((cnp->cn_flags & NOEXECCHECK) != 0) {
6987 cnp->cn_flags &= ~NOEXECCHECK;
6988 return (0);
6989 }
6990
6991 return (VOP_ACCESS(vp, VEXEC, cnp->cn_cred, curthread));
6992 }
6993
6994 /*
6995 * Do not use this variant unless you have means other than the hold count
6996 * to prevent the vnode from getting freed.
6997 */
6998 void
6999 vn_seqc_write_begin_locked(struct vnode *vp)
7000 {
7001
7002 ASSERT_VI_LOCKED(vp, __func__);
7003 VNPASS(vp->v_holdcnt > 0, vp);
7004 VNPASS(vp->v_seqc_users >= 0, vp);
7005 vp->v_seqc_users++;
7006 if (vp->v_seqc_users == 1)
7007 seqc_sleepable_write_begin(&vp->v_seqc);
7008 }
7009
7010 void
7011 vn_seqc_write_begin(struct vnode *vp)
7012 {
7013
7014 VI_LOCK(vp);
7015 vn_seqc_write_begin_locked(vp);
7016 VI_UNLOCK(vp);
7017 }
7018
7019 void
7020 vn_seqc_write_end_locked(struct vnode *vp)
7021 {
7022
7023 ASSERT_VI_LOCKED(vp, __func__);
7024 VNPASS(vp->v_seqc_users > 0, vp);
7025 vp->v_seqc_users--;
7026 if (vp->v_seqc_users == 0)
7027 seqc_sleepable_write_end(&vp->v_seqc);
7028 }
7029
7030 void
7031 vn_seqc_write_end(struct vnode *vp)
7032 {
7033
7034 VI_LOCK(vp);
7035 vn_seqc_write_end_locked(vp);
7036 VI_UNLOCK(vp);
7037 }
7038
7039 /*
7040 * Special case handling for allocating and freeing vnodes.
7041 *
7042 * The counter remains unchanged on free so that a doomed vnode will
7043 * keep testing as in modify as long as it is accessible with SMR.
7044 */
7045 static void
7046 vn_seqc_init(struct vnode *vp)
7047 {
7048
7049 vp->v_seqc = 0;
7050 vp->v_seqc_users = 0;
7051 }
7052
7053 static void
7054 vn_seqc_write_end_free(struct vnode *vp)
7055 {
7056
7057 VNPASS(seqc_in_modify(vp->v_seqc), vp);
7058 VNPASS(vp->v_seqc_users == 1, vp);
7059 }
7060
7061 void
7062 vn_irflag_set_locked(struct vnode *vp, short toset)
7063 {
7064 short flags;
7065
7066 ASSERT_VI_LOCKED(vp, __func__);
7067 flags = vn_irflag_read(vp);
7068 VNASSERT((flags & toset) == 0, vp,
7069 ("%s: some of the passed flags already set (have %d, passed %d)\n",
7070 __func__, flags, toset));
7071 atomic_store_short(&vp->v_irflag, flags | toset);
7072 }
7073
7074 void
7075 vn_irflag_set(struct vnode *vp, short toset)
7076 {
7077
7078 VI_LOCK(vp);
7079 vn_irflag_set_locked(vp, toset);
7080 VI_UNLOCK(vp);
7081 }
7082
7083 void
7084 vn_irflag_set_cond_locked(struct vnode *vp, short toset)
7085 {
7086 short flags;
7087
7088 ASSERT_VI_LOCKED(vp, __func__);
7089 flags = vn_irflag_read(vp);
7090 atomic_store_short(&vp->v_irflag, flags | toset);
7091 }
7092
7093 void
7094 vn_irflag_set_cond(struct vnode *vp, short toset)
7095 {
7096
7097 VI_LOCK(vp);
7098 vn_irflag_set_cond_locked(vp, toset);
7099 VI_UNLOCK(vp);
7100 }
7101
7102 void
7103 vn_irflag_unset_locked(struct vnode *vp, short tounset)
7104 {
7105 short flags;
7106
7107 ASSERT_VI_LOCKED(vp, __func__);
7108 flags = vn_irflag_read(vp);
7109 VNASSERT((flags & tounset) == tounset, vp,
7110 ("%s: some of the passed flags not set (have %d, passed %d)\n",
7111 __func__, flags, tounset));
7112 atomic_store_short(&vp->v_irflag, flags & ~tounset);
7113 }
7114
7115 void
7116 vn_irflag_unset(struct vnode *vp, short tounset)
7117 {
7118
7119 VI_LOCK(vp);
7120 vn_irflag_unset_locked(vp, tounset);
7121 VI_UNLOCK(vp);
7122 }
7123
7124 int
7125 vn_getsize_locked(struct vnode *vp, off_t *size, struct ucred *cred)
7126 {
7127 struct vattr vattr;
7128 int error;
7129
7130 ASSERT_VOP_LOCKED(vp, __func__);
7131 error = VOP_GETATTR(vp, &vattr, cred);
7132 if (__predict_true(error == 0)) {
7133 if (vattr.va_size <= OFF_MAX)
7134 *size = vattr.va_size;
7135 else
7136 error = EFBIG;
7137 }
7138 return (error);
7139 }
7140
7141 int
7142 vn_getsize(struct vnode *vp, off_t *size, struct ucred *cred)
7143 {
7144 int error;
7145
7146 VOP_LOCK(vp, LK_SHARED);
7147 error = vn_getsize_locked(vp, size, cred);
7148 VOP_UNLOCK(vp);
7149 return (error);
7150 }
7151
7152 #ifdef INVARIANTS
7153 void
7154 vn_set_state_validate(struct vnode *vp, enum vstate state)
7155 {
7156
7157 switch (vp->v_state) {
7158 case VSTATE_UNINITIALIZED:
7159 switch (state) {
7160 case VSTATE_CONSTRUCTED:
7161 case VSTATE_DESTROYING:
7162 return;
7163 default:
7164 break;
7165 }
7166 break;
7167 case VSTATE_CONSTRUCTED:
7168 ASSERT_VOP_ELOCKED(vp, __func__);
7169 switch (state) {
7170 case VSTATE_DESTROYING:
7171 return;
7172 default:
7173 break;
7174 }
7175 break;
7176 case VSTATE_DESTROYING:
7177 ASSERT_VOP_ELOCKED(vp, __func__);
7178 switch (state) {
7179 case VSTATE_DEAD:
7180 return;
7181 default:
7182 break;
7183 }
7184 break;
7185 case VSTATE_DEAD:
7186 switch (state) {
7187 case VSTATE_UNINITIALIZED:
7188 return;
7189 default:
7190 break;
7191 }
7192 break;
7193 }
7194
7195 vn_printf(vp, "invalid state transition %d -> %d\n", vp->v_state, state);
7196 panic("invalid state transition %d -> %d\n", vp->v_state, state);
7197 }
7198 #endif
Cache object: 99ea1e3b1fa50c5371a8ec701becce46
|