FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_subr.c
1 /*-
2 * Copyright (c) 1989, 1993
3 * The Regents of the University of California. All rights reserved.
4 * (c) UNIX System Laboratories, Inc.
5 * All or some portions of this file are derived from material licensed
6 * to the University of California by American Telephone and Telegraph
7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
8 * the permission of UNIX System Laboratories, Inc.
9 *
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
12 * are met:
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 4. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
21 *
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * SUCH DAMAGE.
33 *
34 * @(#)vfs_subr.c 8.31 (Berkeley) 5/26/95
35 */
36
37 /*
38 * External virtual filesystem routines
39 */
40
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD: releng/11.2/sys/kern/vfs_subr.c 332755 2018-04-19 06:03:50Z avg $");
43
44 #include "opt_compat.h"
45 #include "opt_ddb.h"
46 #include "opt_watchdog.h"
47
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/bio.h>
51 #include <sys/buf.h>
52 #include <sys/condvar.h>
53 #include <sys/conf.h>
54 #include <sys/counter.h>
55 #include <sys/dirent.h>
56 #include <sys/event.h>
57 #include <sys/eventhandler.h>
58 #include <sys/extattr.h>
59 #include <sys/file.h>
60 #include <sys/fcntl.h>
61 #include <sys/jail.h>
62 #include <sys/kdb.h>
63 #include <sys/kernel.h>
64 #include <sys/kthread.h>
65 #include <sys/lockf.h>
66 #include <sys/malloc.h>
67 #include <sys/mount.h>
68 #include <sys/namei.h>
69 #include <sys/pctrie.h>
70 #include <sys/priv.h>
71 #include <sys/reboot.h>
72 #include <sys/refcount.h>
73 #include <sys/rwlock.h>
74 #include <sys/sched.h>
75 #include <sys/sleepqueue.h>
76 #include <sys/smp.h>
77 #include <sys/stat.h>
78 #include <sys/sysctl.h>
79 #include <sys/syslog.h>
80 #include <sys/vmmeter.h>
81 #include <sys/vnode.h>
82 #include <sys/watchdog.h>
83
84 #include <machine/stdarg.h>
85
86 #include <security/mac/mac_framework.h>
87
88 #include <vm/vm.h>
89 #include <vm/vm_object.h>
90 #include <vm/vm_extern.h>
91 #include <vm/pmap.h>
92 #include <vm/vm_map.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_kern.h>
95 #include <vm/uma.h>
96
97 #ifdef DDB
98 #include <ddb/ddb.h>
99 #endif
100
101 static void delmntque(struct vnode *vp);
102 static int flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo,
103 int slpflag, int slptimeo);
104 static void syncer_shutdown(void *arg, int howto);
105 static int vtryrecycle(struct vnode *vp);
106 static void v_init_counters(struct vnode *);
107 static void v_incr_usecount(struct vnode *);
108 static void v_incr_usecount_locked(struct vnode *);
109 static void v_incr_devcount(struct vnode *);
110 static void v_decr_devcount(struct vnode *);
111 static void vgonel(struct vnode *);
112 static void vfs_knllock(void *arg);
113 static void vfs_knlunlock(void *arg);
114 static void vfs_knl_assert_locked(void *arg);
115 static void vfs_knl_assert_unlocked(void *arg);
116 static void destroy_vpollinfo(struct vpollinfo *vi);
117
118 /*
119 * Number of vnodes in existence. Increased whenever getnewvnode()
120 * allocates a new vnode, decreased in vdropl() for VI_DOOMED vnode.
121 */
122 static unsigned long numvnodes;
123
124 SYSCTL_ULONG(_vfs, OID_AUTO, numvnodes, CTLFLAG_RD, &numvnodes, 0,
125 "Number of vnodes in existence");
126
127 static counter_u64_t vnodes_created;
128 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, vnodes_created, CTLFLAG_RD, &vnodes_created,
129 "Number of vnodes created by getnewvnode");
130
131 /*
132 * Conversion tables for conversion from vnode types to inode formats
133 * and back.
134 */
135 enum vtype iftovt_tab[16] = {
136 VNON, VFIFO, VCHR, VNON, VDIR, VNON, VBLK, VNON,
137 VREG, VNON, VLNK, VNON, VSOCK, VNON, VNON, VBAD,
138 };
139 int vttoif_tab[10] = {
140 0, S_IFREG, S_IFDIR, S_IFBLK, S_IFCHR, S_IFLNK,
141 S_IFSOCK, S_IFIFO, S_IFMT, S_IFMT
142 };
143
144 /*
145 * List of vnodes that are ready for recycling.
146 */
147 static TAILQ_HEAD(freelst, vnode) vnode_free_list;
148
149 /*
150 * "Free" vnode target. Free vnodes are rarely completely free, but are
151 * just ones that are cheap to recycle. Usually they are for files which
152 * have been stat'd but not read; these usually have inode and namecache
153 * data attached to them. This target is the preferred minimum size of a
154 * sub-cache consisting mostly of such files. The system balances the size
155 * of this sub-cache with its complement to try to prevent either from
156 * thrashing while the other is relatively inactive. The targets express
157 * a preference for the best balance.
158 *
159 * "Above" this target there are 2 further targets (watermarks) related
160 * to recyling of free vnodes. In the best-operating case, the cache is
161 * exactly full, the free list has size between vlowat and vhiwat above the
162 * free target, and recycling from it and normal use maintains this state.
163 * Sometimes the free list is below vlowat or even empty, but this state
164 * is even better for immediate use provided the cache is not full.
165 * Otherwise, vnlru_proc() runs to reclaim enough vnodes (usually non-free
166 * ones) to reach one of these states. The watermarks are currently hard-
167 * coded as 4% and 9% of the available space higher. These and the default
168 * of 25% for wantfreevnodes are too large if the memory size is large.
169 * E.g., 9% of 75% of MAXVNODES is more than 566000 vnodes to reclaim
170 * whenever vnlru_proc() becomes active.
171 */
172 static u_long wantfreevnodes;
173 SYSCTL_ULONG(_vfs, OID_AUTO, wantfreevnodes, CTLFLAG_RW,
174 &wantfreevnodes, 0, "Target for minimum number of \"free\" vnodes");
175 static u_long freevnodes;
176 SYSCTL_ULONG(_vfs, OID_AUTO, freevnodes, CTLFLAG_RD,
177 &freevnodes, 0, "Number of \"free\" vnodes");
178
179 static counter_u64_t recycles_count;
180 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, recycles, CTLFLAG_RD, &recycles_count,
181 "Number of vnodes recycled to meet vnode cache targets");
182
183 /*
184 * Various variables used for debugging the new implementation of
185 * reassignbuf().
186 * XXX these are probably of (very) limited utility now.
187 */
188 static int reassignbufcalls;
189 SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW, &reassignbufcalls, 0,
190 "Number of calls to reassignbuf");
191
192 static counter_u64_t free_owe_inact;
193 SYSCTL_COUNTER_U64(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact,
194 "Number of times free vnodes kept on active list due to VFS "
195 "owing inactivation");
196
197 /* To keep more than one thread at a time from running vfs_getnewfsid */
198 static struct mtx mntid_mtx;
199
200 /*
201 * Lock for any access to the following:
202 * vnode_free_list
203 * numvnodes
204 * freevnodes
205 */
206 static struct mtx vnode_free_list_mtx;
207
208 /* Publicly exported FS */
209 struct nfs_public nfs_pub;
210
211 static uma_zone_t buf_trie_zone;
212
213 /* Zone for allocation of new vnodes - used exclusively by getnewvnode() */
214 static uma_zone_t vnode_zone;
215 static uma_zone_t vnodepoll_zone;
216
217 /*
218 * The workitem queue.
219 *
220 * It is useful to delay writes of file data and filesystem metadata
221 * for tens of seconds so that quickly created and deleted files need
222 * not waste disk bandwidth being created and removed. To realize this,
223 * we append vnodes to a "workitem" queue. When running with a soft
224 * updates implementation, most pending metadata dependencies should
225 * not wait for more than a few seconds. Thus, mounted on block devices
226 * are delayed only about a half the time that file data is delayed.
227 * Similarly, directory updates are more critical, so are only delayed
228 * about a third the time that file data is delayed. Thus, there are
229 * SYNCER_MAXDELAY queues that are processed round-robin at a rate of
230 * one each second (driven off the filesystem syncer process). The
231 * syncer_delayno variable indicates the next queue that is to be processed.
232 * Items that need to be processed soon are placed in this queue:
233 *
234 * syncer_workitem_pending[syncer_delayno]
235 *
236 * A delay of fifteen seconds is done by placing the request fifteen
237 * entries later in the queue:
238 *
239 * syncer_workitem_pending[(syncer_delayno + 15) & syncer_mask]
240 *
241 */
242 static int syncer_delayno;
243 static long syncer_mask;
244 LIST_HEAD(synclist, bufobj);
245 static struct synclist *syncer_workitem_pending;
246 /*
247 * The sync_mtx protects:
248 * bo->bo_synclist
249 * sync_vnode_count
250 * syncer_delayno
251 * syncer_state
252 * syncer_workitem_pending
253 * syncer_worklist_len
254 * rushjob
255 */
256 static struct mtx sync_mtx;
257 static struct cv sync_wakeup;
258
259 #define SYNCER_MAXDELAY 32
260 static int syncer_maxdelay = SYNCER_MAXDELAY; /* maximum delay time */
261 static int syncdelay = 30; /* max time to delay syncing data */
262 static int filedelay = 30; /* time to delay syncing files */
263 SYSCTL_INT(_kern, OID_AUTO, filedelay, CTLFLAG_RW, &filedelay, 0,
264 "Time to delay syncing files (in seconds)");
265 static int dirdelay = 29; /* time to delay syncing directories */
266 SYSCTL_INT(_kern, OID_AUTO, dirdelay, CTLFLAG_RW, &dirdelay, 0,
267 "Time to delay syncing directories (in seconds)");
268 static int metadelay = 28; /* time to delay syncing metadata */
269 SYSCTL_INT(_kern, OID_AUTO, metadelay, CTLFLAG_RW, &metadelay, 0,
270 "Time to delay syncing metadata (in seconds)");
271 static int rushjob; /* number of slots to run ASAP */
272 static int stat_rush_requests; /* number of times I/O speeded up */
273 SYSCTL_INT(_debug, OID_AUTO, rush_requests, CTLFLAG_RW, &stat_rush_requests, 0,
274 "Number of times I/O speeded up (rush requests)");
275
276 /*
277 * When shutting down the syncer, run it at four times normal speed.
278 */
279 #define SYNCER_SHUTDOWN_SPEEDUP 4
280 static int sync_vnode_count;
281 static int syncer_worklist_len;
282 static enum { SYNCER_RUNNING, SYNCER_SHUTTING_DOWN, SYNCER_FINAL_DELAY }
283 syncer_state;
284
285 /* Target for maximum number of vnodes. */
286 int desiredvnodes;
287 static int gapvnodes; /* gap between wanted and desired */
288 static int vhiwat; /* enough extras after expansion */
289 static int vlowat; /* minimal extras before expansion */
290 static int vstir; /* nonzero to stir non-free vnodes */
291 static volatile int vsmalltrigger = 8; /* pref to keep if > this many pages */
292
293 static int
294 sysctl_update_desiredvnodes(SYSCTL_HANDLER_ARGS)
295 {
296 int error, old_desiredvnodes;
297
298 old_desiredvnodes = desiredvnodes;
299 if ((error = sysctl_handle_int(oidp, arg1, arg2, req)) != 0)
300 return (error);
301 if (old_desiredvnodes != desiredvnodes) {
302 wantfreevnodes = desiredvnodes / 4;
303 /* XXX locking seems to be incomplete. */
304 vfs_hash_changesize(desiredvnodes);
305 cache_changesize(desiredvnodes);
306 }
307 return (0);
308 }
309
310 SYSCTL_PROC(_kern, KERN_MAXVNODES, maxvnodes,
311 CTLTYPE_INT | CTLFLAG_MPSAFE | CTLFLAG_RW, &desiredvnodes, 0,
312 sysctl_update_desiredvnodes, "I", "Target for maximum number of vnodes");
313 SYSCTL_ULONG(_kern, OID_AUTO, minvnodes, CTLFLAG_RW,
314 &wantfreevnodes, 0, "Old name for vfs.wantfreevnodes (legacy)");
315 static int vnlru_nowhere;
316 SYSCTL_INT(_debug, OID_AUTO, vnlru_nowhere, CTLFLAG_RW,
317 &vnlru_nowhere, 0, "Number of times the vnlru process ran without success");
318
319 /* Shift count for (uintptr_t)vp to initialize vp->v_hash. */
320 static int vnsz2log;
321
322 /*
323 * Support for the bufobj clean & dirty pctrie.
324 */
325 static void *
326 buf_trie_alloc(struct pctrie *ptree)
327 {
328
329 return uma_zalloc(buf_trie_zone, M_NOWAIT);
330 }
331
332 static void
333 buf_trie_free(struct pctrie *ptree, void *node)
334 {
335
336 uma_zfree(buf_trie_zone, node);
337 }
338 PCTRIE_DEFINE(BUF, buf, b_lblkno, buf_trie_alloc, buf_trie_free);
339
340 /*
341 * Initialize the vnode management data structures.
342 *
343 * Reevaluate the following cap on the number of vnodes after the physical
344 * memory size exceeds 512GB. In the limit, as the physical memory size
345 * grows, the ratio of the memory size in KB to vnodes approaches 64:1.
346 */
347 #ifndef MAXVNODES_MAX
348 #define MAXVNODES_MAX (512 * 1024 * 1024 / 64) /* 8M */
349 #endif
350
351 /*
352 * Initialize a vnode as it first enters the zone.
353 */
354 static int
355 vnode_init(void *mem, int size, int flags)
356 {
357 struct vnode *vp;
358 struct bufobj *bo;
359
360 vp = mem;
361 bzero(vp, size);
362 /*
363 * Setup locks.
364 */
365 vp->v_vnlock = &vp->v_lock;
366 mtx_init(&vp->v_interlock, "vnode interlock", NULL, MTX_DEF);
367 /*
368 * By default, don't allow shared locks unless filesystems opt-in.
369 */
370 lockinit(vp->v_vnlock, PVFS, "vnode", VLKTIMEOUT,
371 LK_NOSHARE | LK_IS_VNODE);
372 /*
373 * Initialize bufobj.
374 */
375 bo = &vp->v_bufobj;
376 bo->__bo_vnode = vp;
377 rw_init(BO_LOCKPTR(bo), "bufobj interlock");
378 bo->bo_private = vp;
379 TAILQ_INIT(&bo->bo_clean.bv_hd);
380 TAILQ_INIT(&bo->bo_dirty.bv_hd);
381 /*
382 * Initialize namecache.
383 */
384 LIST_INIT(&vp->v_cache_src);
385 TAILQ_INIT(&vp->v_cache_dst);
386 /*
387 * Initialize rangelocks.
388 */
389 rangelock_init(&vp->v_rl);
390 return (0);
391 }
392
393 /*
394 * Free a vnode when it is cleared from the zone.
395 */
396 static void
397 vnode_fini(void *mem, int size)
398 {
399 struct vnode *vp;
400 struct bufobj *bo;
401
402 vp = mem;
403 rangelock_destroy(&vp->v_rl);
404 lockdestroy(vp->v_vnlock);
405 mtx_destroy(&vp->v_interlock);
406 bo = &vp->v_bufobj;
407 rw_destroy(BO_LOCKPTR(bo));
408 }
409
410 /*
411 * Provide the size of NFS nclnode and NFS fh for calculation of the
412 * vnode memory consumption. The size is specified directly to
413 * eliminate dependency on NFS-private header.
414 *
415 * Other filesystems may use bigger or smaller (like UFS and ZFS)
416 * private inode data, but the NFS-based estimation is ample enough.
417 * Still, we care about differences in the size between 64- and 32-bit
418 * platforms.
419 *
420 * Namecache structure size is heuristically
421 * sizeof(struct namecache_ts) + CACHE_PATH_CUTOFF + 1.
422 */
423 #ifdef _LP64
424 #define NFS_NCLNODE_SZ (528 + 64)
425 #define NC_SZ 148
426 #else
427 #define NFS_NCLNODE_SZ (360 + 32)
428 #define NC_SZ 92
429 #endif
430
431 static void
432 vntblinit(void *dummy __unused)
433 {
434 u_int i;
435 int physvnodes, virtvnodes;
436
437 /*
438 * Desiredvnodes is a function of the physical memory size and the
439 * kernel's heap size. Generally speaking, it scales with the
440 * physical memory size. The ratio of desiredvnodes to the physical
441 * memory size is 1:16 until desiredvnodes exceeds 98,304.
442 * Thereafter, the
443 * marginal ratio of desiredvnodes to the physical memory size is
444 * 1:64. However, desiredvnodes is limited by the kernel's heap
445 * size. The memory required by desiredvnodes vnodes and vm objects
446 * must not exceed 1/10th of the kernel's heap size.
447 */
448 physvnodes = maxproc + pgtok(vm_cnt.v_page_count) / 64 +
449 3 * min(98304 * 16, pgtok(vm_cnt.v_page_count)) / 64;
450 virtvnodes = vm_kmem_size / (10 * (sizeof(struct vm_object) +
451 sizeof(struct vnode) + NC_SZ * ncsizefactor + NFS_NCLNODE_SZ));
452 desiredvnodes = min(physvnodes, virtvnodes);
453 if (desiredvnodes > MAXVNODES_MAX) {
454 if (bootverbose)
455 printf("Reducing kern.maxvnodes %d -> %d\n",
456 desiredvnodes, MAXVNODES_MAX);
457 desiredvnodes = MAXVNODES_MAX;
458 }
459 wantfreevnodes = desiredvnodes / 4;
460 mtx_init(&mntid_mtx, "mntid", NULL, MTX_DEF);
461 TAILQ_INIT(&vnode_free_list);
462 mtx_init(&vnode_free_list_mtx, "vnode_free_list", NULL, MTX_DEF);
463 vnode_zone = uma_zcreate("VNODE", sizeof (struct vnode), NULL, NULL,
464 vnode_init, vnode_fini, UMA_ALIGN_PTR, 0);
465 vnodepoll_zone = uma_zcreate("VNODEPOLL", sizeof (struct vpollinfo),
466 NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0);
467 /*
468 * Preallocate enough nodes to support one-per buf so that
469 * we can not fail an insert. reassignbuf() callers can not
470 * tolerate the insertion failure.
471 */
472 buf_trie_zone = uma_zcreate("BUF TRIE", pctrie_node_size(),
473 NULL, NULL, pctrie_zone_init, NULL, UMA_ALIGN_PTR,
474 UMA_ZONE_NOFREE | UMA_ZONE_VM);
475 uma_prealloc(buf_trie_zone, nbuf);
476
477 vnodes_created = counter_u64_alloc(M_WAITOK);
478 recycles_count = counter_u64_alloc(M_WAITOK);
479 free_owe_inact = counter_u64_alloc(M_WAITOK);
480
481 /*
482 * Initialize the filesystem syncer.
483 */
484 syncer_workitem_pending = hashinit(syncer_maxdelay, M_VNODE,
485 &syncer_mask);
486 syncer_maxdelay = syncer_mask + 1;
487 mtx_init(&sync_mtx, "Syncer mtx", NULL, MTX_DEF);
488 cv_init(&sync_wakeup, "syncer");
489 for (i = 1; i <= sizeof(struct vnode); i <<= 1)
490 vnsz2log++;
491 vnsz2log--;
492 }
493 SYSINIT(vfs, SI_SUB_VFS, SI_ORDER_FIRST, vntblinit, NULL);
494
495
496 /*
497 * Mark a mount point as busy. Used to synchronize access and to delay
498 * unmounting. Eventually, mountlist_mtx is not released on failure.
499 *
500 * vfs_busy() is a custom lock, it can block the caller.
501 * vfs_busy() only sleeps if the unmount is active on the mount point.
502 * For a mountpoint mp, vfs_busy-enforced lock is before lock of any
503 * vnode belonging to mp.
504 *
505 * Lookup uses vfs_busy() to traverse mount points.
506 * root fs var fs
507 * / vnode lock A / vnode lock (/var) D
508 * /var vnode lock B /log vnode lock(/var/log) E
509 * vfs_busy lock C vfs_busy lock F
510 *
511 * Within each file system, the lock order is C->A->B and F->D->E.
512 *
513 * When traversing across mounts, the system follows that lock order:
514 *
515 * C->A->B
516 * |
517 * +->F->D->E
518 *
519 * The lookup() process for namei("/var") illustrates the process:
520 * VOP_LOOKUP() obtains B while A is held
521 * vfs_busy() obtains a shared lock on F while A and B are held
522 * vput() releases lock on B
523 * vput() releases lock on A
524 * VFS_ROOT() obtains lock on D while shared lock on F is held
525 * vfs_unbusy() releases shared lock on F
526 * vn_lock() obtains lock on deadfs vnode vp_crossmp instead of A.
527 * Attempt to lock A (instead of vp_crossmp) while D is held would
528 * violate the global order, causing deadlocks.
529 *
530 * dounmount() locks B while F is drained.
531 */
532 int
533 vfs_busy(struct mount *mp, int flags)
534 {
535
536 MPASS((flags & ~MBF_MASK) == 0);
537 CTR3(KTR_VFS, "%s: mp %p with flags %d", __func__, mp, flags);
538
539 MNT_ILOCK(mp);
540 MNT_REF(mp);
541 /*
542 * If mount point is currently being unmounted, sleep until the
543 * mount point fate is decided. If thread doing the unmounting fails,
544 * it will clear MNTK_UNMOUNT flag before waking us up, indicating
545 * that this mount point has survived the unmount attempt and vfs_busy
546 * should retry. Otherwise the unmounter thread will set MNTK_REFEXPIRE
547 * flag in addition to MNTK_UNMOUNT, indicating that mount point is
548 * about to be really destroyed. vfs_busy needs to release its
549 * reference on the mount point in this case and return with ENOENT,
550 * telling the caller that mount mount it tried to busy is no longer
551 * valid.
552 */
553 while (mp->mnt_kern_flag & MNTK_UNMOUNT) {
554 if (flags & MBF_NOWAIT || mp->mnt_kern_flag & MNTK_REFEXPIRE) {
555 MNT_REL(mp);
556 MNT_IUNLOCK(mp);
557 CTR1(KTR_VFS, "%s: failed busying before sleeping",
558 __func__);
559 return (ENOENT);
560 }
561 if (flags & MBF_MNTLSTLOCK)
562 mtx_unlock(&mountlist_mtx);
563 mp->mnt_kern_flag |= MNTK_MWAIT;
564 msleep(mp, MNT_MTX(mp), PVFS | PDROP, "vfs_busy", 0);
565 if (flags & MBF_MNTLSTLOCK)
566 mtx_lock(&mountlist_mtx);
567 MNT_ILOCK(mp);
568 }
569 if (flags & MBF_MNTLSTLOCK)
570 mtx_unlock(&mountlist_mtx);
571 mp->mnt_lockref++;
572 MNT_IUNLOCK(mp);
573 return (0);
574 }
575
576 /*
577 * Free a busy filesystem.
578 */
579 void
580 vfs_unbusy(struct mount *mp)
581 {
582
583 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
584 MNT_ILOCK(mp);
585 MNT_REL(mp);
586 KASSERT(mp->mnt_lockref > 0, ("negative mnt_lockref"));
587 mp->mnt_lockref--;
588 if (mp->mnt_lockref == 0 && (mp->mnt_kern_flag & MNTK_DRAINING) != 0) {
589 MPASS(mp->mnt_kern_flag & MNTK_UNMOUNT);
590 CTR1(KTR_VFS, "%s: waking up waiters", __func__);
591 mp->mnt_kern_flag &= ~MNTK_DRAINING;
592 wakeup(&mp->mnt_lockref);
593 }
594 MNT_IUNLOCK(mp);
595 }
596
597 /*
598 * Lookup a mount point by filesystem identifier.
599 */
600 struct mount *
601 vfs_getvfs(fsid_t *fsid)
602 {
603 struct mount *mp;
604
605 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
606 mtx_lock(&mountlist_mtx);
607 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
608 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
609 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
610 vfs_ref(mp);
611 mtx_unlock(&mountlist_mtx);
612 return (mp);
613 }
614 }
615 mtx_unlock(&mountlist_mtx);
616 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
617 return ((struct mount *) 0);
618 }
619
620 /*
621 * Lookup a mount point by filesystem identifier, busying it before
622 * returning.
623 *
624 * To avoid congestion on mountlist_mtx, implement simple direct-mapped
625 * cache for popular filesystem identifiers. The cache is lockess, using
626 * the fact that struct mount's are never freed. In worst case we may
627 * get pointer to unmounted or even different filesystem, so we have to
628 * check what we got, and go slow way if so.
629 */
630 struct mount *
631 vfs_busyfs(fsid_t *fsid)
632 {
633 #define FSID_CACHE_SIZE 256
634 typedef struct mount * volatile vmp_t;
635 static vmp_t cache[FSID_CACHE_SIZE];
636 struct mount *mp;
637 int error;
638 uint32_t hash;
639
640 CTR2(KTR_VFS, "%s: fsid %p", __func__, fsid);
641 hash = fsid->val[0] ^ fsid->val[1];
642 hash = (hash >> 16 ^ hash) & (FSID_CACHE_SIZE - 1);
643 mp = cache[hash];
644 if (mp == NULL ||
645 mp->mnt_stat.f_fsid.val[0] != fsid->val[0] ||
646 mp->mnt_stat.f_fsid.val[1] != fsid->val[1])
647 goto slow;
648 if (vfs_busy(mp, 0) != 0) {
649 cache[hash] = NULL;
650 goto slow;
651 }
652 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
653 mp->mnt_stat.f_fsid.val[1] == fsid->val[1])
654 return (mp);
655 else
656 vfs_unbusy(mp);
657
658 slow:
659 mtx_lock(&mountlist_mtx);
660 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
661 if (mp->mnt_stat.f_fsid.val[0] == fsid->val[0] &&
662 mp->mnt_stat.f_fsid.val[1] == fsid->val[1]) {
663 error = vfs_busy(mp, MBF_MNTLSTLOCK);
664 if (error) {
665 cache[hash] = NULL;
666 mtx_unlock(&mountlist_mtx);
667 return (NULL);
668 }
669 cache[hash] = mp;
670 return (mp);
671 }
672 }
673 CTR2(KTR_VFS, "%s: lookup failed for %p id", __func__, fsid);
674 mtx_unlock(&mountlist_mtx);
675 return ((struct mount *) 0);
676 }
677
678 /*
679 * Check if a user can access privileged mount options.
680 */
681 int
682 vfs_suser(struct mount *mp, struct thread *td)
683 {
684 int error;
685
686 /*
687 * If the thread is jailed, but this is not a jail-friendly file
688 * system, deny immediately.
689 */
690 if (!(mp->mnt_vfc->vfc_flags & VFCF_JAIL) && jailed(td->td_ucred))
691 return (EPERM);
692
693 /*
694 * If the file system was mounted outside the jail of the calling
695 * thread, deny immediately.
696 */
697 if (prison_check(td->td_ucred, mp->mnt_cred) != 0)
698 return (EPERM);
699
700 /*
701 * If file system supports delegated administration, we don't check
702 * for the PRIV_VFS_MOUNT_OWNER privilege - it will be better verified
703 * by the file system itself.
704 * If this is not the user that did original mount, we check for
705 * the PRIV_VFS_MOUNT_OWNER privilege.
706 */
707 if (!(mp->mnt_vfc->vfc_flags & VFCF_DELEGADMIN) &&
708 mp->mnt_cred->cr_uid != td->td_ucred->cr_uid) {
709 if ((error = priv_check(td, PRIV_VFS_MOUNT_OWNER)) != 0)
710 return (error);
711 }
712 return (0);
713 }
714
715 /*
716 * Get a new unique fsid. Try to make its val[0] unique, since this value
717 * will be used to create fake device numbers for stat(). Also try (but
718 * not so hard) make its val[0] unique mod 2^16, since some emulators only
719 * support 16-bit device numbers. We end up with unique val[0]'s for the
720 * first 2^16 calls and unique val[0]'s mod 2^16 for the first 2^8 calls.
721 *
722 * Keep in mind that several mounts may be running in parallel. Starting
723 * the search one past where the previous search terminated is both a
724 * micro-optimization and a defense against returning the same fsid to
725 * different mounts.
726 */
727 void
728 vfs_getnewfsid(struct mount *mp)
729 {
730 static uint16_t mntid_base;
731 struct mount *nmp;
732 fsid_t tfsid;
733 int mtype;
734
735 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
736 mtx_lock(&mntid_mtx);
737 mtype = mp->mnt_vfc->vfc_typenum;
738 tfsid.val[1] = mtype;
739 mtype = (mtype & 0xFF) << 24;
740 for (;;) {
741 tfsid.val[0] = makedev(255,
742 mtype | ((mntid_base & 0xFF00) << 8) | (mntid_base & 0xFF));
743 mntid_base++;
744 if ((nmp = vfs_getvfs(&tfsid)) == NULL)
745 break;
746 vfs_rel(nmp);
747 }
748 mp->mnt_stat.f_fsid.val[0] = tfsid.val[0];
749 mp->mnt_stat.f_fsid.val[1] = tfsid.val[1];
750 mtx_unlock(&mntid_mtx);
751 }
752
753 /*
754 * Knob to control the precision of file timestamps:
755 *
756 * 0 = seconds only; nanoseconds zeroed.
757 * 1 = seconds and nanoseconds, accurate within 1/HZ.
758 * 2 = seconds and nanoseconds, truncated to microseconds.
759 * >=3 = seconds and nanoseconds, maximum precision.
760 */
761 enum { TSP_SEC, TSP_HZ, TSP_USEC, TSP_NSEC };
762
763 static int timestamp_precision = TSP_USEC;
764 SYSCTL_INT(_vfs, OID_AUTO, timestamp_precision, CTLFLAG_RW,
765 ×tamp_precision, 0, "File timestamp precision (0: seconds, "
766 "1: sec + ns accurate to 1/HZ, 2: sec + ns truncated to us, "
767 "3+: sec + ns (max. precision))");
768
769 /*
770 * Get a current timestamp.
771 */
772 void
773 vfs_timestamp(struct timespec *tsp)
774 {
775 struct timeval tv;
776
777 switch (timestamp_precision) {
778 case TSP_SEC:
779 tsp->tv_sec = time_second;
780 tsp->tv_nsec = 0;
781 break;
782 case TSP_HZ:
783 getnanotime(tsp);
784 break;
785 case TSP_USEC:
786 microtime(&tv);
787 TIMEVAL_TO_TIMESPEC(&tv, tsp);
788 break;
789 case TSP_NSEC:
790 default:
791 nanotime(tsp);
792 break;
793 }
794 }
795
796 /*
797 * Set vnode attributes to VNOVAL
798 */
799 void
800 vattr_null(struct vattr *vap)
801 {
802
803 vap->va_type = VNON;
804 vap->va_size = VNOVAL;
805 vap->va_bytes = VNOVAL;
806 vap->va_mode = VNOVAL;
807 vap->va_nlink = VNOVAL;
808 vap->va_uid = VNOVAL;
809 vap->va_gid = VNOVAL;
810 vap->va_fsid = VNOVAL;
811 vap->va_fileid = VNOVAL;
812 vap->va_blocksize = VNOVAL;
813 vap->va_rdev = VNOVAL;
814 vap->va_atime.tv_sec = VNOVAL;
815 vap->va_atime.tv_nsec = VNOVAL;
816 vap->va_mtime.tv_sec = VNOVAL;
817 vap->va_mtime.tv_nsec = VNOVAL;
818 vap->va_ctime.tv_sec = VNOVAL;
819 vap->va_ctime.tv_nsec = VNOVAL;
820 vap->va_birthtime.tv_sec = VNOVAL;
821 vap->va_birthtime.tv_nsec = VNOVAL;
822 vap->va_flags = VNOVAL;
823 vap->va_gen = VNOVAL;
824 vap->va_vaflags = 0;
825 }
826
827 /*
828 * This routine is called when we have too many vnodes. It attempts
829 * to free <count> vnodes and will potentially free vnodes that still
830 * have VM backing store (VM backing store is typically the cause
831 * of a vnode blowout so we want to do this). Therefore, this operation
832 * is not considered cheap.
833 *
834 * A number of conditions may prevent a vnode from being reclaimed.
835 * the buffer cache may have references on the vnode, a directory
836 * vnode may still have references due to the namei cache representing
837 * underlying files, or the vnode may be in active use. It is not
838 * desirable to reuse such vnodes. These conditions may cause the
839 * number of vnodes to reach some minimum value regardless of what
840 * you set kern.maxvnodes to. Do not set kern.maxvnodes too low.
841 */
842 static int
843 vlrureclaim(struct mount *mp, int reclaim_nc_src, int trigger)
844 {
845 struct vnode *vp;
846 int count, done, target;
847
848 done = 0;
849 vn_start_write(NULL, &mp, V_WAIT);
850 MNT_ILOCK(mp);
851 count = mp->mnt_nvnodelistsize;
852 target = count * (int64_t)gapvnodes / imax(desiredvnodes, 1);
853 target = target / 10 + 1;
854 while (count != 0 && done < target) {
855 vp = TAILQ_FIRST(&mp->mnt_nvnodelist);
856 while (vp != NULL && vp->v_type == VMARKER)
857 vp = TAILQ_NEXT(vp, v_nmntvnodes);
858 if (vp == NULL)
859 break;
860 /*
861 * XXX LRU is completely broken for non-free vnodes. First
862 * by calling here in mountpoint order, then by moving
863 * unselected vnodes to the end here, and most grossly by
864 * removing the vlruvp() function that was supposed to
865 * maintain the order. (This function was born broken
866 * since syncer problems prevented it doing anything.) The
867 * order is closer to LRC (C = Created).
868 *
869 * LRU reclaiming of vnodes seems to have last worked in
870 * FreeBSD-3 where LRU wasn't mentioned under any spelling.
871 * Then there was no hold count, and inactive vnodes were
872 * simply put on the free list in LRU order. The separate
873 * lists also break LRU. We prefer to reclaim from the
874 * free list for technical reasons. This tends to thrash
875 * the free list to keep very unrecently used held vnodes.
876 * The problem is mitigated by keeping the free list large.
877 */
878 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
879 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
880 --count;
881 if (!VI_TRYLOCK(vp))
882 goto next_iter;
883 /*
884 * If it's been deconstructed already, it's still
885 * referenced, or it exceeds the trigger, skip it.
886 * Also skip free vnodes. We are trying to make space
887 * to expand the free list, not reduce it.
888 */
889 if (vp->v_usecount ||
890 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
891 ((vp->v_iflag & VI_FREE) != 0) ||
892 (vp->v_iflag & VI_DOOMED) != 0 || (vp->v_object != NULL &&
893 vp->v_object->resident_page_count > trigger)) {
894 VI_UNLOCK(vp);
895 goto next_iter;
896 }
897 MNT_IUNLOCK(mp);
898 vholdl(vp);
899 if (VOP_LOCK(vp, LK_INTERLOCK|LK_EXCLUSIVE|LK_NOWAIT)) {
900 vdrop(vp);
901 goto next_iter_mntunlocked;
902 }
903 VI_LOCK(vp);
904 /*
905 * v_usecount may have been bumped after VOP_LOCK() dropped
906 * the vnode interlock and before it was locked again.
907 *
908 * It is not necessary to recheck VI_DOOMED because it can
909 * only be set by another thread that holds both the vnode
910 * lock and vnode interlock. If another thread has the
911 * vnode lock before we get to VOP_LOCK() and obtains the
912 * vnode interlock after VOP_LOCK() drops the vnode
913 * interlock, the other thread will be unable to drop the
914 * vnode lock before our VOP_LOCK() call fails.
915 */
916 if (vp->v_usecount ||
917 (!reclaim_nc_src && !LIST_EMPTY(&vp->v_cache_src)) ||
918 (vp->v_iflag & VI_FREE) != 0 ||
919 (vp->v_object != NULL &&
920 vp->v_object->resident_page_count > trigger)) {
921 VOP_UNLOCK(vp, LK_INTERLOCK);
922 vdrop(vp);
923 goto next_iter_mntunlocked;
924 }
925 KASSERT((vp->v_iflag & VI_DOOMED) == 0,
926 ("VI_DOOMED unexpectedly detected in vlrureclaim()"));
927 counter_u64_add(recycles_count, 1);
928 vgonel(vp);
929 VOP_UNLOCK(vp, 0);
930 vdropl(vp);
931 done++;
932 next_iter_mntunlocked:
933 if (!should_yield())
934 goto relock_mnt;
935 goto yield;
936 next_iter:
937 if (!should_yield())
938 continue;
939 MNT_IUNLOCK(mp);
940 yield:
941 kern_yield(PRI_USER);
942 relock_mnt:
943 MNT_ILOCK(mp);
944 }
945 MNT_IUNLOCK(mp);
946 vn_finished_write(mp);
947 return done;
948 }
949
950 static int max_vnlru_free = 10000; /* limit on vnode free requests per call */
951 SYSCTL_INT(_debug, OID_AUTO, max_vnlru_free, CTLFLAG_RW, &max_vnlru_free,
952 0,
953 "limit on vnode free requests per call to the vnlru_free routine");
954
955 /*
956 * Attempt to reduce the free list by the requested amount.
957 */
958 static void
959 vnlru_free_locked(int count, struct vfsops *mnt_op)
960 {
961 struct vnode *vp;
962 struct mount *mp;
963
964 mtx_assert(&vnode_free_list_mtx, MA_OWNED);
965 if (count > max_vnlru_free)
966 count = max_vnlru_free;
967 for (; count > 0; count--) {
968 vp = TAILQ_FIRST(&vnode_free_list);
969 /*
970 * The list can be modified while the free_list_mtx
971 * has been dropped and vp could be NULL here.
972 */
973 if (!vp)
974 break;
975 VNASSERT(vp->v_op != NULL, vp,
976 ("vnlru_free: vnode already reclaimed."));
977 KASSERT((vp->v_iflag & VI_FREE) != 0,
978 ("Removing vnode not on freelist"));
979 KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
980 ("Mangling active vnode"));
981 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
982
983 /*
984 * Don't recycle if our vnode is from different type
985 * of mount point. Note that mp is type-safe, the
986 * check does not reach unmapped address even if
987 * vnode is reclaimed.
988 * Don't recycle if we can't get the interlock without
989 * blocking.
990 */
991 if ((mnt_op != NULL && (mp = vp->v_mount) != NULL &&
992 mp->mnt_op != mnt_op) || !VI_TRYLOCK(vp)) {
993 TAILQ_INSERT_TAIL(&vnode_free_list, vp, v_actfreelist);
994 continue;
995 }
996 VNASSERT((vp->v_iflag & VI_FREE) != 0 && vp->v_holdcnt == 0,
997 vp, ("vp inconsistent on freelist"));
998
999 /*
1000 * The clear of VI_FREE prevents activation of the
1001 * vnode. There is no sense in putting the vnode on
1002 * the mount point active list, only to remove it
1003 * later during recycling. Inline the relevant part
1004 * of vholdl(), to avoid triggering assertions or
1005 * activating.
1006 */
1007 freevnodes--;
1008 vp->v_iflag &= ~VI_FREE;
1009 refcount_acquire(&vp->v_holdcnt);
1010
1011 mtx_unlock(&vnode_free_list_mtx);
1012 VI_UNLOCK(vp);
1013 vtryrecycle(vp);
1014 /*
1015 * If the recycled succeeded this vdrop will actually free
1016 * the vnode. If not it will simply place it back on
1017 * the free list.
1018 */
1019 vdrop(vp);
1020 mtx_lock(&vnode_free_list_mtx);
1021 }
1022 }
1023
1024 void
1025 vnlru_free(int count, struct vfsops *mnt_op)
1026 {
1027
1028 mtx_lock(&vnode_free_list_mtx);
1029 vnlru_free_locked(count, mnt_op);
1030 mtx_unlock(&vnode_free_list_mtx);
1031 }
1032
1033
1034 /* XXX some names and initialization are bad for limits and watermarks. */
1035 static int
1036 vspace(void)
1037 {
1038 int space;
1039
1040 gapvnodes = imax(desiredvnodes - wantfreevnodes, 100);
1041 vhiwat = gapvnodes / 11; /* 9% -- just under the 10% in vlrureclaim() */
1042 vlowat = vhiwat / 2;
1043 if (numvnodes > desiredvnodes)
1044 return (0);
1045 space = desiredvnodes - numvnodes;
1046 if (freevnodes > wantfreevnodes)
1047 space += freevnodes - wantfreevnodes;
1048 return (space);
1049 }
1050
1051 /*
1052 * Attempt to recycle vnodes in a context that is always safe to block.
1053 * Calling vlrurecycle() from the bowels of filesystem code has some
1054 * interesting deadlock problems.
1055 */
1056 static struct proc *vnlruproc;
1057 static int vnlruproc_sig;
1058
1059 static void
1060 vnlru_proc(void)
1061 {
1062 struct mount *mp, *nmp;
1063 unsigned long ofreevnodes, onumvnodes;
1064 int done, force, reclaim_nc_src, trigger, usevnodes;
1065
1066 EVENTHANDLER_REGISTER(shutdown_pre_sync, kproc_shutdown, vnlruproc,
1067 SHUTDOWN_PRI_FIRST);
1068
1069 force = 0;
1070 for (;;) {
1071 kproc_suspend_check(vnlruproc);
1072 mtx_lock(&vnode_free_list_mtx);
1073 /*
1074 * If numvnodes is too large (due to desiredvnodes being
1075 * adjusted using its sysctl, or emergency growth), first
1076 * try to reduce it by discarding from the free list.
1077 */
1078 if (numvnodes > desiredvnodes && freevnodes > 0)
1079 vnlru_free_locked(ulmin(numvnodes - desiredvnodes,
1080 freevnodes), NULL);
1081 /*
1082 * Sleep if the vnode cache is in a good state. This is
1083 * when it is not over-full and has space for about a 4%
1084 * or 9% expansion (by growing its size or inexcessively
1085 * reducing its free list). Otherwise, try to reclaim
1086 * space for a 10% expansion.
1087 */
1088 if (vstir && force == 0) {
1089 force = 1;
1090 vstir = 0;
1091 }
1092 if (vspace() >= vlowat && force == 0) {
1093 vnlruproc_sig = 0;
1094 wakeup(&vnlruproc_sig);
1095 msleep(vnlruproc, &vnode_free_list_mtx,
1096 PVFS|PDROP, "vlruwt", hz);
1097 continue;
1098 }
1099 mtx_unlock(&vnode_free_list_mtx);
1100 done = 0;
1101 ofreevnodes = freevnodes;
1102 onumvnodes = numvnodes;
1103 /*
1104 * Calculate parameters for recycling. These are the same
1105 * throughout the loop to give some semblance of fairness.
1106 * The trigger point is to avoid recycling vnodes with lots
1107 * of resident pages. We aren't trying to free memory; we
1108 * are trying to recycle or at least free vnodes.
1109 */
1110 if (numvnodes <= desiredvnodes)
1111 usevnodes = numvnodes - freevnodes;
1112 else
1113 usevnodes = numvnodes;
1114 if (usevnodes <= 0)
1115 usevnodes = 1;
1116 /*
1117 * The trigger value is is chosen to give a conservatively
1118 * large value to ensure that it alone doesn't prevent
1119 * making progress. The value can easily be so large that
1120 * it is effectively infinite in some congested and
1121 * misconfigured cases, and this is necessary. Normally
1122 * it is about 8 to 100 (pages), which is quite large.
1123 */
1124 trigger = vm_cnt.v_page_count * 2 / usevnodes;
1125 if (force < 2)
1126 trigger = vsmalltrigger;
1127 reclaim_nc_src = force >= 3;
1128 mtx_lock(&mountlist_mtx);
1129 for (mp = TAILQ_FIRST(&mountlist); mp != NULL; mp = nmp) {
1130 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK)) {
1131 nmp = TAILQ_NEXT(mp, mnt_list);
1132 continue;
1133 }
1134 done += vlrureclaim(mp, reclaim_nc_src, trigger);
1135 mtx_lock(&mountlist_mtx);
1136 nmp = TAILQ_NEXT(mp, mnt_list);
1137 vfs_unbusy(mp);
1138 }
1139 mtx_unlock(&mountlist_mtx);
1140 if (onumvnodes > desiredvnodes && numvnodes <= desiredvnodes)
1141 uma_reclaim();
1142 if (done == 0) {
1143 if (force == 0 || force == 1) {
1144 force = 2;
1145 continue;
1146 }
1147 if (force == 2) {
1148 force = 3;
1149 continue;
1150 }
1151 force = 0;
1152 vnlru_nowhere++;
1153 tsleep(vnlruproc, PPAUSE, "vlrup", hz * 3);
1154 } else
1155 kern_yield(PRI_USER);
1156 /*
1157 * After becoming active to expand above low water, keep
1158 * active until above high water.
1159 */
1160 force = vspace() < vhiwat;
1161 }
1162 }
1163
1164 static struct kproc_desc vnlru_kp = {
1165 "vnlru",
1166 vnlru_proc,
1167 &vnlruproc
1168 };
1169 SYSINIT(vnlru, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start,
1170 &vnlru_kp);
1171
1172 /*
1173 * Routines having to do with the management of the vnode table.
1174 */
1175
1176 /*
1177 * Try to recycle a freed vnode. We abort if anyone picks up a reference
1178 * before we actually vgone(). This function must be called with the vnode
1179 * held to prevent the vnode from being returned to the free list midway
1180 * through vgone().
1181 */
1182 static int
1183 vtryrecycle(struct vnode *vp)
1184 {
1185 struct mount *vnmp;
1186
1187 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
1188 VNASSERT(vp->v_holdcnt, vp,
1189 ("vtryrecycle: Recycling vp %p without a reference.", vp));
1190 /*
1191 * This vnode may found and locked via some other list, if so we
1192 * can't recycle it yet.
1193 */
1194 if (VOP_LOCK(vp, LK_EXCLUSIVE | LK_NOWAIT) != 0) {
1195 CTR2(KTR_VFS,
1196 "%s: impossible to recycle, vp %p lock is already held",
1197 __func__, vp);
1198 return (EWOULDBLOCK);
1199 }
1200 /*
1201 * Don't recycle if its filesystem is being suspended.
1202 */
1203 if (vn_start_write(vp, &vnmp, V_NOWAIT) != 0) {
1204 VOP_UNLOCK(vp, 0);
1205 CTR2(KTR_VFS,
1206 "%s: impossible to recycle, cannot start the write for %p",
1207 __func__, vp);
1208 return (EBUSY);
1209 }
1210 /*
1211 * If we got this far, we need to acquire the interlock and see if
1212 * anyone picked up this vnode from another list. If not, we will
1213 * mark it with DOOMED via vgonel() so that anyone who does find it
1214 * will skip over it.
1215 */
1216 VI_LOCK(vp);
1217 if (vp->v_usecount) {
1218 VOP_UNLOCK(vp, LK_INTERLOCK);
1219 vn_finished_write(vnmp);
1220 CTR2(KTR_VFS,
1221 "%s: impossible to recycle, %p is already referenced",
1222 __func__, vp);
1223 return (EBUSY);
1224 }
1225 if ((vp->v_iflag & VI_DOOMED) == 0) {
1226 counter_u64_add(recycles_count, 1);
1227 vgonel(vp);
1228 }
1229 VOP_UNLOCK(vp, LK_INTERLOCK);
1230 vn_finished_write(vnmp);
1231 return (0);
1232 }
1233
1234 static void
1235 vcheckspace(void)
1236 {
1237
1238 if (vspace() < vlowat && vnlruproc_sig == 0) {
1239 vnlruproc_sig = 1;
1240 wakeup(vnlruproc);
1241 }
1242 }
1243
1244 /*
1245 * Wait if necessary for space for a new vnode.
1246 */
1247 static int
1248 getnewvnode_wait(int suspended)
1249 {
1250
1251 mtx_assert(&vnode_free_list_mtx, MA_OWNED);
1252 if (numvnodes >= desiredvnodes) {
1253 if (suspended) {
1254 /*
1255 * The file system is being suspended. We cannot
1256 * risk a deadlock here, so allow allocation of
1257 * another vnode even if this would give too many.
1258 */
1259 return (0);
1260 }
1261 if (vnlruproc_sig == 0) {
1262 vnlruproc_sig = 1; /* avoid unnecessary wakeups */
1263 wakeup(vnlruproc);
1264 }
1265 msleep(&vnlruproc_sig, &vnode_free_list_mtx, PVFS,
1266 "vlruwk", hz);
1267 }
1268 /* Post-adjust like the pre-adjust in getnewvnode(). */
1269 if (numvnodes + 1 > desiredvnodes && freevnodes > 1)
1270 vnlru_free_locked(1, NULL);
1271 return (numvnodes >= desiredvnodes ? ENFILE : 0);
1272 }
1273
1274 /*
1275 * This hack is fragile, and probably not needed any more now that the
1276 * watermark handling works.
1277 */
1278 void
1279 getnewvnode_reserve(u_int count)
1280 {
1281 struct thread *td;
1282
1283 /* Pre-adjust like the pre-adjust in getnewvnode(), with any count. */
1284 /* XXX no longer so quick, but this part is not racy. */
1285 mtx_lock(&vnode_free_list_mtx);
1286 if (numvnodes + count > desiredvnodes && freevnodes > wantfreevnodes)
1287 vnlru_free_locked(ulmin(numvnodes + count - desiredvnodes,
1288 freevnodes - wantfreevnodes), NULL);
1289 mtx_unlock(&vnode_free_list_mtx);
1290
1291 td = curthread;
1292 /* First try to be quick and racy. */
1293 if (atomic_fetchadd_long(&numvnodes, count) + count <= desiredvnodes) {
1294 td->td_vp_reserv += count;
1295 vcheckspace(); /* XXX no longer so quick, but more racy */
1296 return;
1297 } else
1298 atomic_subtract_long(&numvnodes, count);
1299
1300 mtx_lock(&vnode_free_list_mtx);
1301 while (count > 0) {
1302 if (getnewvnode_wait(0) == 0) {
1303 count--;
1304 td->td_vp_reserv++;
1305 atomic_add_long(&numvnodes, 1);
1306 }
1307 }
1308 vcheckspace();
1309 mtx_unlock(&vnode_free_list_mtx);
1310 }
1311
1312 /*
1313 * This hack is fragile, especially if desiredvnodes or wantvnodes are
1314 * misconfgured or changed significantly. Reducing desiredvnodes below
1315 * the reserved amount should cause bizarre behaviour like reducing it
1316 * below the number of active vnodes -- the system will try to reduce
1317 * numvnodes to match, but should fail, so the subtraction below should
1318 * not overflow.
1319 */
1320 void
1321 getnewvnode_drop_reserve(void)
1322 {
1323 struct thread *td;
1324
1325 td = curthread;
1326 atomic_subtract_long(&numvnodes, td->td_vp_reserv);
1327 td->td_vp_reserv = 0;
1328 }
1329
1330 /*
1331 * Return the next vnode from the free list.
1332 */
1333 int
1334 getnewvnode(const char *tag, struct mount *mp, struct vop_vector *vops,
1335 struct vnode **vpp)
1336 {
1337 struct vnode *vp;
1338 struct thread *td;
1339 struct lock_object *lo;
1340 static int cyclecount;
1341 int error;
1342
1343 CTR3(KTR_VFS, "%s: mp %p with tag %s", __func__, mp, tag);
1344 vp = NULL;
1345 td = curthread;
1346 if (td->td_vp_reserv > 0) {
1347 td->td_vp_reserv -= 1;
1348 goto alloc;
1349 }
1350 mtx_lock(&vnode_free_list_mtx);
1351 if (numvnodes < desiredvnodes)
1352 cyclecount = 0;
1353 else if (cyclecount++ >= freevnodes) {
1354 cyclecount = 0;
1355 vstir = 1;
1356 }
1357 /*
1358 * Grow the vnode cache if it will not be above its target max
1359 * after growing. Otherwise, if the free list is nonempty, try
1360 * to reclaim 1 item from it before growing the cache (possibly
1361 * above its target max if the reclamation failed or is delayed).
1362 * Otherwise, wait for some space. In all cases, schedule
1363 * vnlru_proc() if we are getting short of space. The watermarks
1364 * should be chosen so that we never wait or even reclaim from
1365 * the free list to below its target minimum.
1366 */
1367 if (numvnodes + 1 <= desiredvnodes)
1368 ;
1369 else if (freevnodes > 0)
1370 vnlru_free_locked(1, NULL);
1371 else {
1372 error = getnewvnode_wait(mp != NULL && (mp->mnt_kern_flag &
1373 MNTK_SUSPEND));
1374 #if 0 /* XXX Not all VFS_VGET/ffs_vget callers check returns. */
1375 if (error != 0) {
1376 mtx_unlock(&vnode_free_list_mtx);
1377 return (error);
1378 }
1379 #endif
1380 }
1381 vcheckspace();
1382 atomic_add_long(&numvnodes, 1);
1383 mtx_unlock(&vnode_free_list_mtx);
1384 alloc:
1385 counter_u64_add(vnodes_created, 1);
1386 vp = (struct vnode *) uma_zalloc(vnode_zone, M_WAITOK);
1387 /*
1388 * Locks are given the generic name "vnode" when created.
1389 * Follow the historic practice of using the filesystem
1390 * name when they allocated, e.g., "zfs", "ufs", "nfs, etc.
1391 *
1392 * Locks live in a witness group keyed on their name. Thus,
1393 * when a lock is renamed, it must also move from the witness
1394 * group of its old name to the witness group of its new name.
1395 *
1396 * The change only needs to be made when the vnode moves
1397 * from one filesystem type to another. We ensure that each
1398 * filesystem use a single static name pointer for its tag so
1399 * that we can compare pointers rather than doing a strcmp().
1400 */
1401 lo = &vp->v_vnlock->lock_object;
1402 if (lo->lo_name != tag) {
1403 lo->lo_name = tag;
1404 WITNESS_DESTROY(lo);
1405 WITNESS_INIT(lo, tag);
1406 }
1407 /*
1408 * By default, don't allow shared locks unless filesystems opt-in.
1409 */
1410 vp->v_vnlock->lock_object.lo_flags |= LK_NOSHARE;
1411 /*
1412 * Finalize various vnode identity bits.
1413 */
1414 KASSERT(vp->v_object == NULL, ("stale v_object %p", vp));
1415 KASSERT(vp->v_lockf == NULL, ("stale v_lockf %p", vp));
1416 KASSERT(vp->v_pollinfo == NULL, ("stale v_pollinfo %p", vp));
1417 vp->v_type = VNON;
1418 vp->v_tag = tag;
1419 vp->v_op = vops;
1420 v_init_counters(vp);
1421 vp->v_bufobj.bo_ops = &buf_ops_bio;
1422 #ifdef DIAGNOSTIC
1423 if (mp == NULL && vops != &dead_vnodeops)
1424 printf("NULL mp in getnewvnode(9), tag %s\n", tag);
1425 #endif
1426 #ifdef MAC
1427 mac_vnode_init(vp);
1428 if (mp != NULL && (mp->mnt_flag & MNT_MULTILABEL) == 0)
1429 mac_vnode_associate_singlelabel(mp, vp);
1430 #endif
1431 if (mp != NULL) {
1432 vp->v_bufobj.bo_bsize = mp->mnt_stat.f_iosize;
1433 if ((mp->mnt_kern_flag & MNTK_NOKNOTE) != 0)
1434 vp->v_vflag |= VV_NOKNOTE;
1435 }
1436
1437 /*
1438 * For the filesystems which do not use vfs_hash_insert(),
1439 * still initialize v_hash to have vfs_hash_index() useful.
1440 * E.g., nullfs uses vfs_hash_index() on the lower vnode for
1441 * its own hashing.
1442 */
1443 vp->v_hash = (uintptr_t)vp >> vnsz2log;
1444
1445 *vpp = vp;
1446 return (0);
1447 }
1448
1449 /*
1450 * Delete from old mount point vnode list, if on one.
1451 */
1452 static void
1453 delmntque(struct vnode *vp)
1454 {
1455 struct mount *mp;
1456 int active;
1457
1458 mp = vp->v_mount;
1459 if (mp == NULL)
1460 return;
1461 MNT_ILOCK(mp);
1462 VI_LOCK(vp);
1463 KASSERT(mp->mnt_activevnodelistsize <= mp->mnt_nvnodelistsize,
1464 ("Active vnode list size %d > Vnode list size %d",
1465 mp->mnt_activevnodelistsize, mp->mnt_nvnodelistsize));
1466 active = vp->v_iflag & VI_ACTIVE;
1467 vp->v_iflag &= ~VI_ACTIVE;
1468 if (active) {
1469 mtx_lock(&vnode_free_list_mtx);
1470 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
1471 mp->mnt_activevnodelistsize--;
1472 mtx_unlock(&vnode_free_list_mtx);
1473 }
1474 vp->v_mount = NULL;
1475 VI_UNLOCK(vp);
1476 VNASSERT(mp->mnt_nvnodelistsize > 0, vp,
1477 ("bad mount point vnode list size"));
1478 TAILQ_REMOVE(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1479 mp->mnt_nvnodelistsize--;
1480 MNT_REL(mp);
1481 MNT_IUNLOCK(mp);
1482 }
1483
1484 static void
1485 insmntque_stddtr(struct vnode *vp, void *dtr_arg)
1486 {
1487
1488 vp->v_data = NULL;
1489 vp->v_op = &dead_vnodeops;
1490 vgone(vp);
1491 vput(vp);
1492 }
1493
1494 /*
1495 * Insert into list of vnodes for the new mount point, if available.
1496 */
1497 int
1498 insmntque1(struct vnode *vp, struct mount *mp,
1499 void (*dtr)(struct vnode *, void *), void *dtr_arg)
1500 {
1501
1502 KASSERT(vp->v_mount == NULL,
1503 ("insmntque: vnode already on per mount vnode list"));
1504 VNASSERT(mp != NULL, vp, ("Don't call insmntque(foo, NULL)"));
1505 ASSERT_VOP_ELOCKED(vp, "insmntque: non-locked vp");
1506
1507 /*
1508 * We acquire the vnode interlock early to ensure that the
1509 * vnode cannot be recycled by another process releasing a
1510 * holdcnt on it before we get it on both the vnode list
1511 * and the active vnode list. The mount mutex protects only
1512 * manipulation of the vnode list and the vnode freelist
1513 * mutex protects only manipulation of the active vnode list.
1514 * Hence the need to hold the vnode interlock throughout.
1515 */
1516 MNT_ILOCK(mp);
1517 VI_LOCK(vp);
1518 if (((mp->mnt_kern_flag & MNTK_NOINSMNTQ) != 0 &&
1519 ((mp->mnt_kern_flag & MNTK_UNMOUNTF) != 0 ||
1520 mp->mnt_nvnodelistsize == 0)) &&
1521 (vp->v_vflag & VV_FORCEINSMQ) == 0) {
1522 VI_UNLOCK(vp);
1523 MNT_IUNLOCK(mp);
1524 if (dtr != NULL)
1525 dtr(vp, dtr_arg);
1526 return (EBUSY);
1527 }
1528 vp->v_mount = mp;
1529 MNT_REF(mp);
1530 TAILQ_INSERT_TAIL(&mp->mnt_nvnodelist, vp, v_nmntvnodes);
1531 VNASSERT(mp->mnt_nvnodelistsize >= 0, vp,
1532 ("neg mount point vnode list size"));
1533 mp->mnt_nvnodelistsize++;
1534 KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
1535 ("Activating already active vnode"));
1536 vp->v_iflag |= VI_ACTIVE;
1537 mtx_lock(&vnode_free_list_mtx);
1538 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
1539 mp->mnt_activevnodelistsize++;
1540 mtx_unlock(&vnode_free_list_mtx);
1541 VI_UNLOCK(vp);
1542 MNT_IUNLOCK(mp);
1543 return (0);
1544 }
1545
1546 int
1547 insmntque(struct vnode *vp, struct mount *mp)
1548 {
1549
1550 return (insmntque1(vp, mp, insmntque_stddtr, NULL));
1551 }
1552
1553 /*
1554 * Flush out and invalidate all buffers associated with a bufobj
1555 * Called with the underlying object locked.
1556 */
1557 int
1558 bufobj_invalbuf(struct bufobj *bo, int flags, int slpflag, int slptimeo)
1559 {
1560 int error;
1561
1562 BO_LOCK(bo);
1563 if (flags & V_SAVE) {
1564 error = bufobj_wwait(bo, slpflag, slptimeo);
1565 if (error) {
1566 BO_UNLOCK(bo);
1567 return (error);
1568 }
1569 if (bo->bo_dirty.bv_cnt > 0) {
1570 BO_UNLOCK(bo);
1571 if ((error = BO_SYNC(bo, MNT_WAIT)) != 0)
1572 return (error);
1573 /*
1574 * XXX We could save a lock/unlock if this was only
1575 * enabled under INVARIANTS
1576 */
1577 BO_LOCK(bo);
1578 if (bo->bo_numoutput > 0 || bo->bo_dirty.bv_cnt > 0)
1579 panic("vinvalbuf: dirty bufs");
1580 }
1581 }
1582 /*
1583 * If you alter this loop please notice that interlock is dropped and
1584 * reacquired in flushbuflist. Special care is needed to ensure that
1585 * no race conditions occur from this.
1586 */
1587 do {
1588 error = flushbuflist(&bo->bo_clean,
1589 flags, bo, slpflag, slptimeo);
1590 if (error == 0 && !(flags & V_CLEANONLY))
1591 error = flushbuflist(&bo->bo_dirty,
1592 flags, bo, slpflag, slptimeo);
1593 if (error != 0 && error != EAGAIN) {
1594 BO_UNLOCK(bo);
1595 return (error);
1596 }
1597 } while (error != 0);
1598
1599 /*
1600 * Wait for I/O to complete. XXX needs cleaning up. The vnode can
1601 * have write I/O in-progress but if there is a VM object then the
1602 * VM object can also have read-I/O in-progress.
1603 */
1604 do {
1605 bufobj_wwait(bo, 0, 0);
1606 if ((flags & V_VMIO) == 0) {
1607 BO_UNLOCK(bo);
1608 if (bo->bo_object != NULL) {
1609 VM_OBJECT_WLOCK(bo->bo_object);
1610 vm_object_pip_wait(bo->bo_object, "bovlbx");
1611 VM_OBJECT_WUNLOCK(bo->bo_object);
1612 }
1613 BO_LOCK(bo);
1614 }
1615 } while (bo->bo_numoutput > 0);
1616 BO_UNLOCK(bo);
1617
1618 /*
1619 * Destroy the copy in the VM cache, too.
1620 */
1621 if (bo->bo_object != NULL &&
1622 (flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0) {
1623 VM_OBJECT_WLOCK(bo->bo_object);
1624 vm_object_page_remove(bo->bo_object, 0, 0, (flags & V_SAVE) ?
1625 OBJPR_CLEANONLY : 0);
1626 VM_OBJECT_WUNLOCK(bo->bo_object);
1627 }
1628
1629 #ifdef INVARIANTS
1630 BO_LOCK(bo);
1631 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO |
1632 V_ALLOWCLEAN)) == 0 && (bo->bo_dirty.bv_cnt > 0 ||
1633 bo->bo_clean.bv_cnt > 0))
1634 panic("vinvalbuf: flush failed");
1635 if ((flags & (V_ALT | V_NORMAL | V_CLEANONLY | V_VMIO)) == 0 &&
1636 bo->bo_dirty.bv_cnt > 0)
1637 panic("vinvalbuf: flush dirty failed");
1638 BO_UNLOCK(bo);
1639 #endif
1640 return (0);
1641 }
1642
1643 /*
1644 * Flush out and invalidate all buffers associated with a vnode.
1645 * Called with the underlying object locked.
1646 */
1647 int
1648 vinvalbuf(struct vnode *vp, int flags, int slpflag, int slptimeo)
1649 {
1650
1651 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
1652 ASSERT_VOP_LOCKED(vp, "vinvalbuf");
1653 if (vp->v_object != NULL && vp->v_object->handle != vp)
1654 return (0);
1655 return (bufobj_invalbuf(&vp->v_bufobj, flags, slpflag, slptimeo));
1656 }
1657
1658 /*
1659 * Flush out buffers on the specified list.
1660 *
1661 */
1662 static int
1663 flushbuflist(struct bufv *bufv, int flags, struct bufobj *bo, int slpflag,
1664 int slptimeo)
1665 {
1666 struct buf *bp, *nbp;
1667 int retval, error;
1668 daddr_t lblkno;
1669 b_xflags_t xflags;
1670
1671 ASSERT_BO_WLOCKED(bo);
1672
1673 retval = 0;
1674 TAILQ_FOREACH_SAFE(bp, &bufv->bv_hd, b_bobufs, nbp) {
1675 if (((flags & V_NORMAL) && (bp->b_xflags & BX_ALTDATA)) ||
1676 ((flags & V_ALT) && (bp->b_xflags & BX_ALTDATA) == 0)) {
1677 continue;
1678 }
1679 if (nbp != NULL) {
1680 lblkno = nbp->b_lblkno;
1681 xflags = nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN);
1682 }
1683 retval = EAGAIN;
1684 error = BUF_TIMELOCK(bp,
1685 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK, BO_LOCKPTR(bo),
1686 "flushbuf", slpflag, slptimeo);
1687 if (error) {
1688 BO_LOCK(bo);
1689 return (error != ENOLCK ? error : EAGAIN);
1690 }
1691 KASSERT(bp->b_bufobj == bo,
1692 ("bp %p wrong b_bufobj %p should be %p",
1693 bp, bp->b_bufobj, bo));
1694 /*
1695 * XXX Since there are no node locks for NFS, I
1696 * believe there is a slight chance that a delayed
1697 * write will occur while sleeping just above, so
1698 * check for it.
1699 */
1700 if (((bp->b_flags & (B_DELWRI | B_INVAL)) == B_DELWRI) &&
1701 (flags & V_SAVE)) {
1702 bremfree(bp);
1703 bp->b_flags |= B_ASYNC;
1704 bwrite(bp);
1705 BO_LOCK(bo);
1706 return (EAGAIN); /* XXX: why not loop ? */
1707 }
1708 bremfree(bp);
1709 bp->b_flags |= (B_INVAL | B_RELBUF);
1710 bp->b_flags &= ~B_ASYNC;
1711 brelse(bp);
1712 BO_LOCK(bo);
1713 if (nbp == NULL)
1714 break;
1715 nbp = gbincore(bo, lblkno);
1716 if (nbp == NULL || (nbp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1717 != xflags)
1718 break; /* nbp invalid */
1719 }
1720 return (retval);
1721 }
1722
1723 int
1724 bnoreuselist(struct bufv *bufv, struct bufobj *bo, daddr_t startn, daddr_t endn)
1725 {
1726 struct buf *bp;
1727 int error;
1728 daddr_t lblkno;
1729
1730 ASSERT_BO_LOCKED(bo);
1731
1732 for (lblkno = startn;;) {
1733 again:
1734 bp = BUF_PCTRIE_LOOKUP_GE(&bufv->bv_root, lblkno);
1735 if (bp == NULL || bp->b_lblkno >= endn ||
1736 bp->b_lblkno < startn)
1737 break;
1738 error = BUF_TIMELOCK(bp, LK_EXCLUSIVE | LK_SLEEPFAIL |
1739 LK_INTERLOCK, BO_LOCKPTR(bo), "brlsfl", 0, 0);
1740 if (error != 0) {
1741 BO_RLOCK(bo);
1742 if (error == ENOLCK)
1743 goto again;
1744 return (error);
1745 }
1746 KASSERT(bp->b_bufobj == bo,
1747 ("bp %p wrong b_bufobj %p should be %p",
1748 bp, bp->b_bufobj, bo));
1749 lblkno = bp->b_lblkno + 1;
1750 if ((bp->b_flags & B_MANAGED) == 0)
1751 bremfree(bp);
1752 bp->b_flags |= B_RELBUF;
1753 /*
1754 * In the VMIO case, use the B_NOREUSE flag to hint that the
1755 * pages backing each buffer in the range are unlikely to be
1756 * reused. Dirty buffers will have the hint applied once
1757 * they've been written.
1758 */
1759 if (bp->b_vp->v_object != NULL)
1760 bp->b_flags |= B_NOREUSE;
1761 brelse(bp);
1762 BO_RLOCK(bo);
1763 }
1764 return (0);
1765 }
1766
1767 /*
1768 * Truncate a file's buffer and pages to a specified length. This
1769 * is in lieu of the old vinvalbuf mechanism, which performed unneeded
1770 * sync activity.
1771 */
1772 int
1773 vtruncbuf(struct vnode *vp, struct ucred *cred, off_t length, int blksize)
1774 {
1775 struct buf *bp, *nbp;
1776 int anyfreed;
1777 int trunclbn;
1778 struct bufobj *bo;
1779
1780 CTR5(KTR_VFS, "%s: vp %p with cred %p and block %d:%ju", __func__,
1781 vp, cred, blksize, (uintmax_t)length);
1782
1783 /*
1784 * Round up to the *next* lbn.
1785 */
1786 trunclbn = howmany(length, blksize);
1787
1788 ASSERT_VOP_LOCKED(vp, "vtruncbuf");
1789 restart:
1790 bo = &vp->v_bufobj;
1791 BO_LOCK(bo);
1792 anyfreed = 1;
1793 for (;anyfreed;) {
1794 anyfreed = 0;
1795 TAILQ_FOREACH_SAFE(bp, &bo->bo_clean.bv_hd, b_bobufs, nbp) {
1796 if (bp->b_lblkno < trunclbn)
1797 continue;
1798 if (BUF_LOCK(bp,
1799 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1800 BO_LOCKPTR(bo)) == ENOLCK)
1801 goto restart;
1802
1803 bremfree(bp);
1804 bp->b_flags |= (B_INVAL | B_RELBUF);
1805 bp->b_flags &= ~B_ASYNC;
1806 brelse(bp);
1807 anyfreed = 1;
1808
1809 BO_LOCK(bo);
1810 if (nbp != NULL &&
1811 (((nbp->b_xflags & BX_VNCLEAN) == 0) ||
1812 (nbp->b_vp != vp) ||
1813 (nbp->b_flags & B_DELWRI))) {
1814 BO_UNLOCK(bo);
1815 goto restart;
1816 }
1817 }
1818
1819 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1820 if (bp->b_lblkno < trunclbn)
1821 continue;
1822 if (BUF_LOCK(bp,
1823 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1824 BO_LOCKPTR(bo)) == ENOLCK)
1825 goto restart;
1826 bremfree(bp);
1827 bp->b_flags |= (B_INVAL | B_RELBUF);
1828 bp->b_flags &= ~B_ASYNC;
1829 brelse(bp);
1830 anyfreed = 1;
1831
1832 BO_LOCK(bo);
1833 if (nbp != NULL &&
1834 (((nbp->b_xflags & BX_VNDIRTY) == 0) ||
1835 (nbp->b_vp != vp) ||
1836 (nbp->b_flags & B_DELWRI) == 0)) {
1837 BO_UNLOCK(bo);
1838 goto restart;
1839 }
1840 }
1841 }
1842
1843 if (length > 0) {
1844 restartsync:
1845 TAILQ_FOREACH_SAFE(bp, &bo->bo_dirty.bv_hd, b_bobufs, nbp) {
1846 if (bp->b_lblkno > 0)
1847 continue;
1848 /*
1849 * Since we hold the vnode lock this should only
1850 * fail if we're racing with the buf daemon.
1851 */
1852 if (BUF_LOCK(bp,
1853 LK_EXCLUSIVE | LK_SLEEPFAIL | LK_INTERLOCK,
1854 BO_LOCKPTR(bo)) == ENOLCK) {
1855 goto restart;
1856 }
1857 VNASSERT((bp->b_flags & B_DELWRI), vp,
1858 ("buf(%p) on dirty queue without DELWRI", bp));
1859
1860 bremfree(bp);
1861 bawrite(bp);
1862 BO_LOCK(bo);
1863 goto restartsync;
1864 }
1865 }
1866
1867 bufobj_wwait(bo, 0, 0);
1868 BO_UNLOCK(bo);
1869 vnode_pager_setsize(vp, length);
1870
1871 return (0);
1872 }
1873
1874 static void
1875 buf_vlist_remove(struct buf *bp)
1876 {
1877 struct bufv *bv;
1878
1879 KASSERT(bp->b_bufobj != NULL, ("No b_bufobj %p", bp));
1880 ASSERT_BO_WLOCKED(bp->b_bufobj);
1881 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) !=
1882 (BX_VNDIRTY|BX_VNCLEAN),
1883 ("buf_vlist_remove: Buf %p is on two lists", bp));
1884 if (bp->b_xflags & BX_VNDIRTY)
1885 bv = &bp->b_bufobj->bo_dirty;
1886 else
1887 bv = &bp->b_bufobj->bo_clean;
1888 BUF_PCTRIE_REMOVE(&bv->bv_root, bp->b_lblkno);
1889 TAILQ_REMOVE(&bv->bv_hd, bp, b_bobufs);
1890 bv->bv_cnt--;
1891 bp->b_xflags &= ~(BX_VNDIRTY | BX_VNCLEAN);
1892 }
1893
1894 /*
1895 * Add the buffer to the sorted clean or dirty block list.
1896 *
1897 * NOTE: xflags is passed as a constant, optimizing this inline function!
1898 */
1899 static void
1900 buf_vlist_add(struct buf *bp, struct bufobj *bo, b_xflags_t xflags)
1901 {
1902 struct bufv *bv;
1903 struct buf *n;
1904 int error;
1905
1906 ASSERT_BO_WLOCKED(bo);
1907 KASSERT((xflags & BX_VNDIRTY) == 0 || (bo->bo_flag & BO_DEAD) == 0,
1908 ("dead bo %p", bo));
1909 KASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0,
1910 ("buf_vlist_add: Buf %p has existing xflags %d", bp, bp->b_xflags));
1911 bp->b_xflags |= xflags;
1912 if (xflags & BX_VNDIRTY)
1913 bv = &bo->bo_dirty;
1914 else
1915 bv = &bo->bo_clean;
1916
1917 /*
1918 * Keep the list ordered. Optimize empty list insertion. Assume
1919 * we tend to grow at the tail so lookup_le should usually be cheaper
1920 * than _ge.
1921 */
1922 if (bv->bv_cnt == 0 ||
1923 bp->b_lblkno > TAILQ_LAST(&bv->bv_hd, buflists)->b_lblkno)
1924 TAILQ_INSERT_TAIL(&bv->bv_hd, bp, b_bobufs);
1925 else if ((n = BUF_PCTRIE_LOOKUP_LE(&bv->bv_root, bp->b_lblkno)) == NULL)
1926 TAILQ_INSERT_HEAD(&bv->bv_hd, bp, b_bobufs);
1927 else
1928 TAILQ_INSERT_AFTER(&bv->bv_hd, n, bp, b_bobufs);
1929 error = BUF_PCTRIE_INSERT(&bv->bv_root, bp);
1930 if (error)
1931 panic("buf_vlist_add: Preallocated nodes insufficient.");
1932 bv->bv_cnt++;
1933 }
1934
1935 /*
1936 * Look up a buffer using the buffer tries.
1937 */
1938 struct buf *
1939 gbincore(struct bufobj *bo, daddr_t lblkno)
1940 {
1941 struct buf *bp;
1942
1943 ASSERT_BO_LOCKED(bo);
1944 bp = BUF_PCTRIE_LOOKUP(&bo->bo_clean.bv_root, lblkno);
1945 if (bp != NULL)
1946 return (bp);
1947 return BUF_PCTRIE_LOOKUP(&bo->bo_dirty.bv_root, lblkno);
1948 }
1949
1950 /*
1951 * Associate a buffer with a vnode.
1952 */
1953 void
1954 bgetvp(struct vnode *vp, struct buf *bp)
1955 {
1956 struct bufobj *bo;
1957
1958 bo = &vp->v_bufobj;
1959 ASSERT_BO_WLOCKED(bo);
1960 VNASSERT(bp->b_vp == NULL, bp->b_vp, ("bgetvp: not free"));
1961
1962 CTR3(KTR_BUF, "bgetvp(%p) vp %p flags %X", bp, vp, bp->b_flags);
1963 VNASSERT((bp->b_xflags & (BX_VNDIRTY|BX_VNCLEAN)) == 0, vp,
1964 ("bgetvp: bp already attached! %p", bp));
1965
1966 vhold(vp);
1967 bp->b_vp = vp;
1968 bp->b_bufobj = bo;
1969 /*
1970 * Insert onto list for new vnode.
1971 */
1972 buf_vlist_add(bp, bo, BX_VNCLEAN);
1973 }
1974
1975 /*
1976 * Disassociate a buffer from a vnode.
1977 */
1978 void
1979 brelvp(struct buf *bp)
1980 {
1981 struct bufobj *bo;
1982 struct vnode *vp;
1983
1984 CTR3(KTR_BUF, "brelvp(%p) vp %p flags %X", bp, bp->b_vp, bp->b_flags);
1985 KASSERT(bp->b_vp != NULL, ("brelvp: NULL"));
1986
1987 /*
1988 * Delete from old vnode list, if on one.
1989 */
1990 vp = bp->b_vp; /* XXX */
1991 bo = bp->b_bufobj;
1992 BO_LOCK(bo);
1993 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
1994 buf_vlist_remove(bp);
1995 else
1996 panic("brelvp: Buffer %p not on queue.", bp);
1997 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
1998 bo->bo_flag &= ~BO_ONWORKLST;
1999 mtx_lock(&sync_mtx);
2000 LIST_REMOVE(bo, bo_synclist);
2001 syncer_worklist_len--;
2002 mtx_unlock(&sync_mtx);
2003 }
2004 bp->b_vp = NULL;
2005 bp->b_bufobj = NULL;
2006 BO_UNLOCK(bo);
2007 vdrop(vp);
2008 }
2009
2010 /*
2011 * Add an item to the syncer work queue.
2012 */
2013 static void
2014 vn_syncer_add_to_worklist(struct bufobj *bo, int delay)
2015 {
2016 int slot;
2017
2018 ASSERT_BO_WLOCKED(bo);
2019
2020 mtx_lock(&sync_mtx);
2021 if (bo->bo_flag & BO_ONWORKLST)
2022 LIST_REMOVE(bo, bo_synclist);
2023 else {
2024 bo->bo_flag |= BO_ONWORKLST;
2025 syncer_worklist_len++;
2026 }
2027
2028 if (delay > syncer_maxdelay - 2)
2029 delay = syncer_maxdelay - 2;
2030 slot = (syncer_delayno + delay) & syncer_mask;
2031
2032 LIST_INSERT_HEAD(&syncer_workitem_pending[slot], bo, bo_synclist);
2033 mtx_unlock(&sync_mtx);
2034 }
2035
2036 static int
2037 sysctl_vfs_worklist_len(SYSCTL_HANDLER_ARGS)
2038 {
2039 int error, len;
2040
2041 mtx_lock(&sync_mtx);
2042 len = syncer_worklist_len - sync_vnode_count;
2043 mtx_unlock(&sync_mtx);
2044 error = SYSCTL_OUT(req, &len, sizeof(len));
2045 return (error);
2046 }
2047
2048 SYSCTL_PROC(_vfs, OID_AUTO, worklist_len, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
2049 sysctl_vfs_worklist_len, "I", "Syncer thread worklist length");
2050
2051 static struct proc *updateproc;
2052 static void sched_sync(void);
2053 static struct kproc_desc up_kp = {
2054 "syncer",
2055 sched_sync,
2056 &updateproc
2057 };
2058 SYSINIT(syncer, SI_SUB_KTHREAD_UPDATE, SI_ORDER_FIRST, kproc_start, &up_kp);
2059
2060 static int
2061 sync_vnode(struct synclist *slp, struct bufobj **bo, struct thread *td)
2062 {
2063 struct vnode *vp;
2064 struct mount *mp;
2065
2066 *bo = LIST_FIRST(slp);
2067 if (*bo == NULL)
2068 return (0);
2069 vp = (*bo)->__bo_vnode; /* XXX */
2070 if (VOP_ISLOCKED(vp) != 0 || VI_TRYLOCK(vp) == 0)
2071 return (1);
2072 /*
2073 * We use vhold in case the vnode does not
2074 * successfully sync. vhold prevents the vnode from
2075 * going away when we unlock the sync_mtx so that
2076 * we can acquire the vnode interlock.
2077 */
2078 vholdl(vp);
2079 mtx_unlock(&sync_mtx);
2080 VI_UNLOCK(vp);
2081 if (vn_start_write(vp, &mp, V_NOWAIT) != 0) {
2082 vdrop(vp);
2083 mtx_lock(&sync_mtx);
2084 return (*bo == LIST_FIRST(slp));
2085 }
2086 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
2087 (void) VOP_FSYNC(vp, MNT_LAZY, td);
2088 VOP_UNLOCK(vp, 0);
2089 vn_finished_write(mp);
2090 BO_LOCK(*bo);
2091 if (((*bo)->bo_flag & BO_ONWORKLST) != 0) {
2092 /*
2093 * Put us back on the worklist. The worklist
2094 * routine will remove us from our current
2095 * position and then add us back in at a later
2096 * position.
2097 */
2098 vn_syncer_add_to_worklist(*bo, syncdelay);
2099 }
2100 BO_UNLOCK(*bo);
2101 vdrop(vp);
2102 mtx_lock(&sync_mtx);
2103 return (0);
2104 }
2105
2106 static int first_printf = 1;
2107
2108 /*
2109 * System filesystem synchronizer daemon.
2110 */
2111 static void
2112 sched_sync(void)
2113 {
2114 struct synclist *next, *slp;
2115 struct bufobj *bo;
2116 long starttime;
2117 struct thread *td = curthread;
2118 int last_work_seen;
2119 int net_worklist_len;
2120 int syncer_final_iter;
2121 int error;
2122
2123 last_work_seen = 0;
2124 syncer_final_iter = 0;
2125 syncer_state = SYNCER_RUNNING;
2126 starttime = time_uptime;
2127 td->td_pflags |= TDP_NORUNNINGBUF;
2128
2129 EVENTHANDLER_REGISTER(shutdown_pre_sync, syncer_shutdown, td->td_proc,
2130 SHUTDOWN_PRI_LAST);
2131
2132 mtx_lock(&sync_mtx);
2133 for (;;) {
2134 if (syncer_state == SYNCER_FINAL_DELAY &&
2135 syncer_final_iter == 0) {
2136 mtx_unlock(&sync_mtx);
2137 kproc_suspend_check(td->td_proc);
2138 mtx_lock(&sync_mtx);
2139 }
2140 net_worklist_len = syncer_worklist_len - sync_vnode_count;
2141 if (syncer_state != SYNCER_RUNNING &&
2142 starttime != time_uptime) {
2143 if (first_printf) {
2144 printf("\nSyncing disks, vnodes remaining... ");
2145 first_printf = 0;
2146 }
2147 printf("%d ", net_worklist_len);
2148 }
2149 starttime = time_uptime;
2150
2151 /*
2152 * Push files whose dirty time has expired. Be careful
2153 * of interrupt race on slp queue.
2154 *
2155 * Skip over empty worklist slots when shutting down.
2156 */
2157 do {
2158 slp = &syncer_workitem_pending[syncer_delayno];
2159 syncer_delayno += 1;
2160 if (syncer_delayno == syncer_maxdelay)
2161 syncer_delayno = 0;
2162 next = &syncer_workitem_pending[syncer_delayno];
2163 /*
2164 * If the worklist has wrapped since the
2165 * it was emptied of all but syncer vnodes,
2166 * switch to the FINAL_DELAY state and run
2167 * for one more second.
2168 */
2169 if (syncer_state == SYNCER_SHUTTING_DOWN &&
2170 net_worklist_len == 0 &&
2171 last_work_seen == syncer_delayno) {
2172 syncer_state = SYNCER_FINAL_DELAY;
2173 syncer_final_iter = SYNCER_SHUTDOWN_SPEEDUP;
2174 }
2175 } while (syncer_state != SYNCER_RUNNING && LIST_EMPTY(slp) &&
2176 syncer_worklist_len > 0);
2177
2178 /*
2179 * Keep track of the last time there was anything
2180 * on the worklist other than syncer vnodes.
2181 * Return to the SHUTTING_DOWN state if any
2182 * new work appears.
2183 */
2184 if (net_worklist_len > 0 || syncer_state == SYNCER_RUNNING)
2185 last_work_seen = syncer_delayno;
2186 if (net_worklist_len > 0 && syncer_state == SYNCER_FINAL_DELAY)
2187 syncer_state = SYNCER_SHUTTING_DOWN;
2188 while (!LIST_EMPTY(slp)) {
2189 error = sync_vnode(slp, &bo, td);
2190 if (error == 1) {
2191 LIST_REMOVE(bo, bo_synclist);
2192 LIST_INSERT_HEAD(next, bo, bo_synclist);
2193 continue;
2194 }
2195
2196 if (first_printf == 0) {
2197 /*
2198 * Drop the sync mutex, because some watchdog
2199 * drivers need to sleep while patting
2200 */
2201 mtx_unlock(&sync_mtx);
2202 wdog_kern_pat(WD_LASTVAL);
2203 mtx_lock(&sync_mtx);
2204 }
2205
2206 }
2207 if (syncer_state == SYNCER_FINAL_DELAY && syncer_final_iter > 0)
2208 syncer_final_iter--;
2209 /*
2210 * The variable rushjob allows the kernel to speed up the
2211 * processing of the filesystem syncer process. A rushjob
2212 * value of N tells the filesystem syncer to process the next
2213 * N seconds worth of work on its queue ASAP. Currently rushjob
2214 * is used by the soft update code to speed up the filesystem
2215 * syncer process when the incore state is getting so far
2216 * ahead of the disk that the kernel memory pool is being
2217 * threatened with exhaustion.
2218 */
2219 if (rushjob > 0) {
2220 rushjob -= 1;
2221 continue;
2222 }
2223 /*
2224 * Just sleep for a short period of time between
2225 * iterations when shutting down to allow some I/O
2226 * to happen.
2227 *
2228 * If it has taken us less than a second to process the
2229 * current work, then wait. Otherwise start right over
2230 * again. We can still lose time if any single round
2231 * takes more than two seconds, but it does not really
2232 * matter as we are just trying to generally pace the
2233 * filesystem activity.
2234 */
2235 if (syncer_state != SYNCER_RUNNING ||
2236 time_uptime == starttime) {
2237 thread_lock(td);
2238 sched_prio(td, PPAUSE);
2239 thread_unlock(td);
2240 }
2241 if (syncer_state != SYNCER_RUNNING)
2242 cv_timedwait(&sync_wakeup, &sync_mtx,
2243 hz / SYNCER_SHUTDOWN_SPEEDUP);
2244 else if (time_uptime == starttime)
2245 cv_timedwait(&sync_wakeup, &sync_mtx, hz);
2246 }
2247 }
2248
2249 /*
2250 * Request the syncer daemon to speed up its work.
2251 * We never push it to speed up more than half of its
2252 * normal turn time, otherwise it could take over the cpu.
2253 */
2254 int
2255 speedup_syncer(void)
2256 {
2257 int ret = 0;
2258
2259 mtx_lock(&sync_mtx);
2260 if (rushjob < syncdelay / 2) {
2261 rushjob += 1;
2262 stat_rush_requests += 1;
2263 ret = 1;
2264 }
2265 mtx_unlock(&sync_mtx);
2266 cv_broadcast(&sync_wakeup);
2267 return (ret);
2268 }
2269
2270 /*
2271 * Tell the syncer to speed up its work and run though its work
2272 * list several times, then tell it to shut down.
2273 */
2274 static void
2275 syncer_shutdown(void *arg, int howto)
2276 {
2277
2278 if (howto & RB_NOSYNC)
2279 return;
2280 mtx_lock(&sync_mtx);
2281 syncer_state = SYNCER_SHUTTING_DOWN;
2282 rushjob = 0;
2283 mtx_unlock(&sync_mtx);
2284 cv_broadcast(&sync_wakeup);
2285 kproc_shutdown(arg, howto);
2286 }
2287
2288 void
2289 syncer_suspend(void)
2290 {
2291
2292 syncer_shutdown(updateproc, 0);
2293 }
2294
2295 void
2296 syncer_resume(void)
2297 {
2298
2299 mtx_lock(&sync_mtx);
2300 first_printf = 1;
2301 syncer_state = SYNCER_RUNNING;
2302 mtx_unlock(&sync_mtx);
2303 cv_broadcast(&sync_wakeup);
2304 kproc_resume(updateproc);
2305 }
2306
2307 /*
2308 * Reassign a buffer from one vnode to another.
2309 * Used to assign file specific control information
2310 * (indirect blocks) to the vnode to which they belong.
2311 */
2312 void
2313 reassignbuf(struct buf *bp)
2314 {
2315 struct vnode *vp;
2316 struct bufobj *bo;
2317 int delay;
2318 #ifdef INVARIANTS
2319 struct bufv *bv;
2320 #endif
2321
2322 vp = bp->b_vp;
2323 bo = bp->b_bufobj;
2324 ++reassignbufcalls;
2325
2326 CTR3(KTR_BUF, "reassignbuf(%p) vp %p flags %X",
2327 bp, bp->b_vp, bp->b_flags);
2328 /*
2329 * B_PAGING flagged buffers cannot be reassigned because their vp
2330 * is not fully linked in.
2331 */
2332 if (bp->b_flags & B_PAGING)
2333 panic("cannot reassign paging buffer");
2334
2335 /*
2336 * Delete from old vnode list, if on one.
2337 */
2338 BO_LOCK(bo);
2339 if (bp->b_xflags & (BX_VNDIRTY | BX_VNCLEAN))
2340 buf_vlist_remove(bp);
2341 else
2342 panic("reassignbuf: Buffer %p not on queue.", bp);
2343 /*
2344 * If dirty, put on list of dirty buffers; otherwise insert onto list
2345 * of clean buffers.
2346 */
2347 if (bp->b_flags & B_DELWRI) {
2348 if ((bo->bo_flag & BO_ONWORKLST) == 0) {
2349 switch (vp->v_type) {
2350 case VDIR:
2351 delay = dirdelay;
2352 break;
2353 case VCHR:
2354 delay = metadelay;
2355 break;
2356 default:
2357 delay = filedelay;
2358 }
2359 vn_syncer_add_to_worklist(bo, delay);
2360 }
2361 buf_vlist_add(bp, bo, BX_VNDIRTY);
2362 } else {
2363 buf_vlist_add(bp, bo, BX_VNCLEAN);
2364
2365 if ((bo->bo_flag & BO_ONWORKLST) && bo->bo_dirty.bv_cnt == 0) {
2366 mtx_lock(&sync_mtx);
2367 LIST_REMOVE(bo, bo_synclist);
2368 syncer_worklist_len--;
2369 mtx_unlock(&sync_mtx);
2370 bo->bo_flag &= ~BO_ONWORKLST;
2371 }
2372 }
2373 #ifdef INVARIANTS
2374 bv = &bo->bo_clean;
2375 bp = TAILQ_FIRST(&bv->bv_hd);
2376 KASSERT(bp == NULL || bp->b_bufobj == bo,
2377 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2378 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2379 KASSERT(bp == NULL || bp->b_bufobj == bo,
2380 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2381 bv = &bo->bo_dirty;
2382 bp = TAILQ_FIRST(&bv->bv_hd);
2383 KASSERT(bp == NULL || bp->b_bufobj == bo,
2384 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2385 bp = TAILQ_LAST(&bv->bv_hd, buflists);
2386 KASSERT(bp == NULL || bp->b_bufobj == bo,
2387 ("bp %p wrong b_bufobj %p should be %p", bp, bp->b_bufobj, bo));
2388 #endif
2389 BO_UNLOCK(bo);
2390 }
2391
2392 static void
2393 v_init_counters(struct vnode *vp)
2394 {
2395
2396 VNASSERT(vp->v_type == VNON && vp->v_data == NULL && vp->v_iflag == 0,
2397 vp, ("%s called for an initialized vnode", __FUNCTION__));
2398 ASSERT_VI_UNLOCKED(vp, __FUNCTION__);
2399
2400 refcount_init(&vp->v_holdcnt, 1);
2401 refcount_init(&vp->v_usecount, 1);
2402 }
2403
2404 static void
2405 v_incr_usecount_locked(struct vnode *vp)
2406 {
2407
2408 ASSERT_VI_LOCKED(vp, __func__);
2409 if ((vp->v_iflag & VI_OWEINACT) != 0) {
2410 VNASSERT(vp->v_usecount == 0, vp,
2411 ("vnode with usecount and VI_OWEINACT set"));
2412 vp->v_iflag &= ~VI_OWEINACT;
2413 }
2414 refcount_acquire(&vp->v_usecount);
2415 v_incr_devcount(vp);
2416 }
2417
2418 /*
2419 * Increment the use and hold counts on the vnode, taking care to reference
2420 * the driver's usecount if this is a chardev. The _vhold() will remove
2421 * the vnode from the free list if it is presently free.
2422 */
2423 static void
2424 v_incr_usecount(struct vnode *vp)
2425 {
2426
2427 ASSERT_VI_UNLOCKED(vp, __func__);
2428 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2429
2430 if (vp->v_type != VCHR &&
2431 refcount_acquire_if_not_zero(&vp->v_usecount)) {
2432 VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
2433 ("vnode with usecount and VI_OWEINACT set"));
2434 } else {
2435 VI_LOCK(vp);
2436 v_incr_usecount_locked(vp);
2437 VI_UNLOCK(vp);
2438 }
2439 }
2440
2441 /*
2442 * Increment si_usecount of the associated device, if any.
2443 */
2444 static void
2445 v_incr_devcount(struct vnode *vp)
2446 {
2447
2448 ASSERT_VI_LOCKED(vp, __FUNCTION__);
2449 if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2450 dev_lock();
2451 vp->v_rdev->si_usecount++;
2452 dev_unlock();
2453 }
2454 }
2455
2456 /*
2457 * Decrement si_usecount of the associated device, if any.
2458 */
2459 static void
2460 v_decr_devcount(struct vnode *vp)
2461 {
2462
2463 ASSERT_VI_LOCKED(vp, __FUNCTION__);
2464 if (vp->v_type == VCHR && vp->v_rdev != NULL) {
2465 dev_lock();
2466 vp->v_rdev->si_usecount--;
2467 dev_unlock();
2468 }
2469 }
2470
2471 /*
2472 * Grab a particular vnode from the free list, increment its
2473 * reference count and lock it. VI_DOOMED is set if the vnode
2474 * is being destroyed. Only callers who specify LK_RETRY will
2475 * see doomed vnodes. If inactive processing was delayed in
2476 * vput try to do it here.
2477 *
2478 * Notes on lockless counter manipulation:
2479 * _vhold, vputx and other routines make various decisions based
2480 * on either holdcnt or usecount being 0. As long as either counter
2481 * is not transitioning 0->1 nor 1->0, the manipulation can be done
2482 * with atomic operations. Otherwise the interlock is taken covering
2483 * both the atomic and additional actions.
2484 */
2485 int
2486 vget(struct vnode *vp, int flags, struct thread *td)
2487 {
2488 int error, oweinact;
2489
2490 VNASSERT((flags & LK_TYPE_MASK) != 0, vp,
2491 ("vget: invalid lock operation"));
2492
2493 if ((flags & LK_INTERLOCK) != 0)
2494 ASSERT_VI_LOCKED(vp, __func__);
2495 else
2496 ASSERT_VI_UNLOCKED(vp, __func__);
2497 if ((flags & LK_VNHELD) != 0)
2498 VNASSERT((vp->v_holdcnt > 0), vp,
2499 ("vget: LK_VNHELD passed but vnode not held"));
2500
2501 CTR3(KTR_VFS, "%s: vp %p with flags %d", __func__, vp, flags);
2502
2503 if ((flags & LK_VNHELD) == 0)
2504 _vhold(vp, (flags & LK_INTERLOCK) != 0);
2505
2506 if ((error = vn_lock(vp, flags)) != 0) {
2507 vdrop(vp);
2508 CTR2(KTR_VFS, "%s: impossible to lock vnode %p", __func__,
2509 vp);
2510 return (error);
2511 }
2512 if (vp->v_iflag & VI_DOOMED && (flags & LK_RETRY) == 0)
2513 panic("vget: vn_lock failed to return ENOENT\n");
2514 /*
2515 * We don't guarantee that any particular close will
2516 * trigger inactive processing so just make a best effort
2517 * here at preventing a reference to a removed file. If
2518 * we don't succeed no harm is done.
2519 *
2520 * Upgrade our holdcnt to a usecount.
2521 */
2522 if (vp->v_type == VCHR ||
2523 !refcount_acquire_if_not_zero(&vp->v_usecount)) {
2524 VI_LOCK(vp);
2525 if ((vp->v_iflag & VI_OWEINACT) == 0) {
2526 oweinact = 0;
2527 } else {
2528 oweinact = 1;
2529 vp->v_iflag &= ~VI_OWEINACT;
2530 }
2531 refcount_acquire(&vp->v_usecount);
2532 v_incr_devcount(vp);
2533 if (oweinact && VOP_ISLOCKED(vp) == LK_EXCLUSIVE &&
2534 (flags & LK_NOWAIT) == 0)
2535 vinactive(vp, td);
2536 VI_UNLOCK(vp);
2537 }
2538 return (0);
2539 }
2540
2541 /*
2542 * Increase the reference count of a vnode.
2543 */
2544 void
2545 vref(struct vnode *vp)
2546 {
2547
2548 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2549 _vhold(vp, false);
2550 v_incr_usecount(vp);
2551 }
2552
2553 void
2554 vrefl(struct vnode *vp)
2555 {
2556
2557 ASSERT_VI_LOCKED(vp, __func__);
2558 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2559 _vhold(vp, true);
2560 v_incr_usecount_locked(vp);
2561 }
2562
2563 void
2564 vrefact(struct vnode *vp)
2565 {
2566
2567 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2568 if (__predict_false(vp->v_type == VCHR)) {
2569 VNASSERT(vp->v_holdcnt > 0 && vp->v_usecount > 0, vp,
2570 ("%s: wrong ref counts", __func__));
2571 vref(vp);
2572 return;
2573 }
2574 #ifdef INVARIANTS
2575 int old = atomic_fetchadd_int(&vp->v_holdcnt, 1);
2576 VNASSERT(old > 0, vp, ("%s: wrong hold count", __func__));
2577 old = atomic_fetchadd_int(&vp->v_usecount, 1);
2578 VNASSERT(old > 0, vp, ("%s: wrong use count", __func__));
2579 #else
2580 refcount_acquire(&vp->v_holdcnt);
2581 refcount_acquire(&vp->v_usecount);
2582 #endif
2583 }
2584
2585 /*
2586 * Return reference count of a vnode.
2587 *
2588 * The results of this call are only guaranteed when some mechanism is used to
2589 * stop other processes from gaining references to the vnode. This may be the
2590 * case if the caller holds the only reference. This is also useful when stale
2591 * data is acceptable as race conditions may be accounted for by some other
2592 * means.
2593 */
2594 int
2595 vrefcnt(struct vnode *vp)
2596 {
2597
2598 return (vp->v_usecount);
2599 }
2600
2601 #define VPUTX_VRELE 1
2602 #define VPUTX_VPUT 2
2603 #define VPUTX_VUNREF 3
2604
2605 /*
2606 * Decrement the use and hold counts for a vnode.
2607 *
2608 * See an explanation near vget() as to why atomic operation is safe.
2609 */
2610 static void
2611 vputx(struct vnode *vp, int func)
2612 {
2613 int error;
2614
2615 KASSERT(vp != NULL, ("vputx: null vp"));
2616 if (func == VPUTX_VUNREF)
2617 ASSERT_VOP_LOCKED(vp, "vunref");
2618 else if (func == VPUTX_VPUT)
2619 ASSERT_VOP_LOCKED(vp, "vput");
2620 else
2621 KASSERT(func == VPUTX_VRELE, ("vputx: wrong func"));
2622 ASSERT_VI_UNLOCKED(vp, __func__);
2623 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2624
2625 if (vp->v_type != VCHR &&
2626 refcount_release_if_not_last(&vp->v_usecount)) {
2627 if (func == VPUTX_VPUT)
2628 VOP_UNLOCK(vp, 0);
2629 vdrop(vp);
2630 return;
2631 }
2632
2633 VI_LOCK(vp);
2634
2635 /*
2636 * We want to hold the vnode until the inactive finishes to
2637 * prevent vgone() races. We drop the use count here and the
2638 * hold count below when we're done.
2639 */
2640 if (!refcount_release(&vp->v_usecount) ||
2641 (vp->v_iflag & VI_DOINGINACT)) {
2642 if (func == VPUTX_VPUT)
2643 VOP_UNLOCK(vp, 0);
2644 v_decr_devcount(vp);
2645 vdropl(vp);
2646 return;
2647 }
2648
2649 v_decr_devcount(vp);
2650
2651 error = 0;
2652
2653 if (vp->v_usecount != 0) {
2654 vn_printf(vp, "vputx: usecount not zero for vnode ");
2655 panic("vputx: usecount not zero");
2656 }
2657
2658 CTR2(KTR_VFS, "%s: return vnode %p to the freelist", __func__, vp);
2659
2660 /*
2661 * We must call VOP_INACTIVE with the node locked. Mark
2662 * as VI_DOINGINACT to avoid recursion.
2663 */
2664 vp->v_iflag |= VI_OWEINACT;
2665 switch (func) {
2666 case VPUTX_VRELE:
2667 error = vn_lock(vp, LK_EXCLUSIVE | LK_INTERLOCK);
2668 VI_LOCK(vp);
2669 break;
2670 case VPUTX_VPUT:
2671 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2672 error = VOP_LOCK(vp, LK_UPGRADE | LK_INTERLOCK |
2673 LK_NOWAIT);
2674 VI_LOCK(vp);
2675 }
2676 break;
2677 case VPUTX_VUNREF:
2678 if (VOP_ISLOCKED(vp) != LK_EXCLUSIVE) {
2679 error = VOP_LOCK(vp, LK_TRYUPGRADE | LK_INTERLOCK);
2680 VI_LOCK(vp);
2681 }
2682 break;
2683 }
2684 VNASSERT(vp->v_usecount == 0 || (vp->v_iflag & VI_OWEINACT) == 0, vp,
2685 ("vnode with usecount and VI_OWEINACT set"));
2686 if (error == 0) {
2687 if (vp->v_iflag & VI_OWEINACT)
2688 vinactive(vp, curthread);
2689 if (func != VPUTX_VUNREF)
2690 VOP_UNLOCK(vp, 0);
2691 }
2692 vdropl(vp);
2693 }
2694
2695 /*
2696 * Vnode put/release.
2697 * If count drops to zero, call inactive routine and return to freelist.
2698 */
2699 void
2700 vrele(struct vnode *vp)
2701 {
2702
2703 vputx(vp, VPUTX_VRELE);
2704 }
2705
2706 /*
2707 * Release an already locked vnode. This give the same effects as
2708 * unlock+vrele(), but takes less time and avoids releasing and
2709 * re-aquiring the lock (as vrele() acquires the lock internally.)
2710 */
2711 void
2712 vput(struct vnode *vp)
2713 {
2714
2715 vputx(vp, VPUTX_VPUT);
2716 }
2717
2718 /*
2719 * Release an exclusively locked vnode. Do not unlock the vnode lock.
2720 */
2721 void
2722 vunref(struct vnode *vp)
2723 {
2724
2725 vputx(vp, VPUTX_VUNREF);
2726 }
2727
2728 /*
2729 * Increase the hold count and activate if this is the first reference.
2730 */
2731 void
2732 _vhold(struct vnode *vp, bool locked)
2733 {
2734 struct mount *mp;
2735
2736 if (locked)
2737 ASSERT_VI_LOCKED(vp, __func__);
2738 else
2739 ASSERT_VI_UNLOCKED(vp, __func__);
2740 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2741 if (!locked) {
2742 if (refcount_acquire_if_not_zero(&vp->v_holdcnt)) {
2743 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2744 ("_vhold: vnode with holdcnt is free"));
2745 return;
2746 }
2747 VI_LOCK(vp);
2748 }
2749 if ((vp->v_iflag & VI_FREE) == 0) {
2750 refcount_acquire(&vp->v_holdcnt);
2751 if (!locked)
2752 VI_UNLOCK(vp);
2753 return;
2754 }
2755 VNASSERT(vp->v_holdcnt == 0, vp,
2756 ("%s: wrong hold count", __func__));
2757 VNASSERT(vp->v_op != NULL, vp,
2758 ("%s: vnode already reclaimed.", __func__));
2759 /*
2760 * Remove a vnode from the free list, mark it as in use,
2761 * and put it on the active list.
2762 */
2763 mtx_lock(&vnode_free_list_mtx);
2764 TAILQ_REMOVE(&vnode_free_list, vp, v_actfreelist);
2765 freevnodes--;
2766 vp->v_iflag &= ~VI_FREE;
2767 KASSERT((vp->v_iflag & VI_ACTIVE) == 0,
2768 ("Activating already active vnode"));
2769 vp->v_iflag |= VI_ACTIVE;
2770 mp = vp->v_mount;
2771 TAILQ_INSERT_HEAD(&mp->mnt_activevnodelist, vp, v_actfreelist);
2772 mp->mnt_activevnodelistsize++;
2773 mtx_unlock(&vnode_free_list_mtx);
2774 refcount_acquire(&vp->v_holdcnt);
2775 if (!locked)
2776 VI_UNLOCK(vp);
2777 }
2778
2779 /*
2780 * Drop the hold count of the vnode. If this is the last reference to
2781 * the vnode we place it on the free list unless it has been vgone'd
2782 * (marked VI_DOOMED) in which case we will free it.
2783 *
2784 * Because the vnode vm object keeps a hold reference on the vnode if
2785 * there is at least one resident non-cached page, the vnode cannot
2786 * leave the active list without the page cleanup done.
2787 */
2788 void
2789 _vdrop(struct vnode *vp, bool locked)
2790 {
2791 struct bufobj *bo;
2792 struct mount *mp;
2793 int active;
2794
2795 if (locked)
2796 ASSERT_VI_LOCKED(vp, __func__);
2797 else
2798 ASSERT_VI_UNLOCKED(vp, __func__);
2799 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2800 if ((int)vp->v_holdcnt <= 0)
2801 panic("vdrop: holdcnt %d", vp->v_holdcnt);
2802 if (!locked) {
2803 if (refcount_release_if_not_last(&vp->v_holdcnt))
2804 return;
2805 VI_LOCK(vp);
2806 }
2807 if (refcount_release(&vp->v_holdcnt) == 0) {
2808 VI_UNLOCK(vp);
2809 return;
2810 }
2811 if ((vp->v_iflag & VI_DOOMED) == 0) {
2812 /*
2813 * Mark a vnode as free: remove it from its active list
2814 * and put it up for recycling on the freelist.
2815 */
2816 VNASSERT(vp->v_op != NULL, vp,
2817 ("vdropl: vnode already reclaimed."));
2818 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2819 ("vnode already free"));
2820 VNASSERT(vp->v_holdcnt == 0, vp,
2821 ("vdropl: freeing when we shouldn't"));
2822 active = vp->v_iflag & VI_ACTIVE;
2823 if ((vp->v_iflag & VI_OWEINACT) == 0) {
2824 vp->v_iflag &= ~VI_ACTIVE;
2825 mp = vp->v_mount;
2826 mtx_lock(&vnode_free_list_mtx);
2827 if (active) {
2828 TAILQ_REMOVE(&mp->mnt_activevnodelist, vp,
2829 v_actfreelist);
2830 mp->mnt_activevnodelistsize--;
2831 }
2832 TAILQ_INSERT_TAIL(&vnode_free_list, vp,
2833 v_actfreelist);
2834 freevnodes++;
2835 vp->v_iflag |= VI_FREE;
2836 mtx_unlock(&vnode_free_list_mtx);
2837 } else {
2838 counter_u64_add(free_owe_inact, 1);
2839 }
2840 VI_UNLOCK(vp);
2841 return;
2842 }
2843 /*
2844 * The vnode has been marked for destruction, so free it.
2845 *
2846 * The vnode will be returned to the zone where it will
2847 * normally remain until it is needed for another vnode. We
2848 * need to cleanup (or verify that the cleanup has already
2849 * been done) any residual data left from its current use
2850 * so as not to contaminate the freshly allocated vnode.
2851 */
2852 CTR2(KTR_VFS, "%s: destroying the vnode %p", __func__, vp);
2853 atomic_subtract_long(&numvnodes, 1);
2854 bo = &vp->v_bufobj;
2855 VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
2856 ("cleaned vnode still on the free list."));
2857 VNASSERT(vp->v_data == NULL, vp, ("cleaned vnode isn't"));
2858 VNASSERT(vp->v_holdcnt == 0, vp, ("Non-zero hold count"));
2859 VNASSERT(vp->v_usecount == 0, vp, ("Non-zero use count"));
2860 VNASSERT(vp->v_writecount == 0, vp, ("Non-zero write count"));
2861 VNASSERT(bo->bo_numoutput == 0, vp, ("Clean vnode has pending I/O's"));
2862 VNASSERT(bo->bo_clean.bv_cnt == 0, vp, ("cleanbufcnt not 0"));
2863 VNASSERT(pctrie_is_empty(&bo->bo_clean.bv_root), vp,
2864 ("clean blk trie not empty"));
2865 VNASSERT(bo->bo_dirty.bv_cnt == 0, vp, ("dirtybufcnt not 0"));
2866 VNASSERT(pctrie_is_empty(&bo->bo_dirty.bv_root), vp,
2867 ("dirty blk trie not empty"));
2868 VNASSERT(TAILQ_EMPTY(&vp->v_cache_dst), vp, ("vp has namecache dst"));
2869 VNASSERT(LIST_EMPTY(&vp->v_cache_src), vp, ("vp has namecache src"));
2870 VNASSERT(vp->v_cache_dd == NULL, vp, ("vp has namecache for .."));
2871 VNASSERT(TAILQ_EMPTY(&vp->v_rl.rl_waiters), vp,
2872 ("Dangling rangelock waiters"));
2873 VI_UNLOCK(vp);
2874 #ifdef MAC
2875 mac_vnode_destroy(vp);
2876 #endif
2877 if (vp->v_pollinfo != NULL) {
2878 destroy_vpollinfo(vp->v_pollinfo);
2879 vp->v_pollinfo = NULL;
2880 }
2881 #ifdef INVARIANTS
2882 /* XXX Elsewhere we detect an already freed vnode via NULL v_op. */
2883 vp->v_op = NULL;
2884 #endif
2885 bzero(&vp->v_un, sizeof(vp->v_un));
2886 vp->v_lasta = vp->v_clen = vp->v_cstart = vp->v_lastw = 0;
2887 vp->v_iflag = 0;
2888 vp->v_vflag = 0;
2889 bo->bo_flag = 0;
2890 uma_zfree(vnode_zone, vp);
2891 }
2892
2893 /*
2894 * Call VOP_INACTIVE on the vnode and manage the DOINGINACT and OWEINACT
2895 * flags. DOINGINACT prevents us from recursing in calls to vinactive.
2896 * OWEINACT tracks whether a vnode missed a call to inactive due to a
2897 * failed lock upgrade.
2898 */
2899 void
2900 vinactive(struct vnode *vp, struct thread *td)
2901 {
2902 struct vm_object *obj;
2903
2904 ASSERT_VOP_ELOCKED(vp, "vinactive");
2905 ASSERT_VI_LOCKED(vp, "vinactive");
2906 VNASSERT((vp->v_iflag & VI_DOINGINACT) == 0, vp,
2907 ("vinactive: recursed on VI_DOINGINACT"));
2908 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
2909 vp->v_iflag |= VI_DOINGINACT;
2910 vp->v_iflag &= ~VI_OWEINACT;
2911 VI_UNLOCK(vp);
2912 /*
2913 * Before moving off the active list, we must be sure that any
2914 * modified pages are converted into the vnode's dirty
2915 * buffers, since these will no longer be checked once the
2916 * vnode is on the inactive list.
2917 *
2918 * The write-out of the dirty pages is asynchronous. At the
2919 * point that VOP_INACTIVE() is called, there could still be
2920 * pending I/O and dirty pages in the object.
2921 */
2922 if ((obj = vp->v_object) != NULL && (vp->v_vflag & VV_NOSYNC) == 0 &&
2923 (obj->flags & OBJ_MIGHTBEDIRTY) != 0) {
2924 VM_OBJECT_WLOCK(obj);
2925 vm_object_page_clean(obj, 0, 0, 0);
2926 VM_OBJECT_WUNLOCK(obj);
2927 }
2928 VOP_INACTIVE(vp, td);
2929 VI_LOCK(vp);
2930 VNASSERT(vp->v_iflag & VI_DOINGINACT, vp,
2931 ("vinactive: lost VI_DOINGINACT"));
2932 vp->v_iflag &= ~VI_DOINGINACT;
2933 }
2934
2935 /*
2936 * Remove any vnodes in the vnode table belonging to mount point mp.
2937 *
2938 * If FORCECLOSE is not specified, there should not be any active ones,
2939 * return error if any are found (nb: this is a user error, not a
2940 * system error). If FORCECLOSE is specified, detach any active vnodes
2941 * that are found.
2942 *
2943 * If WRITECLOSE is set, only flush out regular file vnodes open for
2944 * writing.
2945 *
2946 * SKIPSYSTEM causes any vnodes marked VV_SYSTEM to be skipped.
2947 *
2948 * `rootrefs' specifies the base reference count for the root vnode
2949 * of this filesystem. The root vnode is considered busy if its
2950 * v_usecount exceeds this value. On a successful return, vflush(, td)
2951 * will call vrele() on the root vnode exactly rootrefs times.
2952 * If the SKIPSYSTEM or WRITECLOSE flags are specified, rootrefs must
2953 * be zero.
2954 */
2955 #ifdef DIAGNOSTIC
2956 static int busyprt = 0; /* print out busy vnodes */
2957 SYSCTL_INT(_debug, OID_AUTO, busyprt, CTLFLAG_RW, &busyprt, 0, "Print out busy vnodes");
2958 #endif
2959
2960 int
2961 vflush(struct mount *mp, int rootrefs, int flags, struct thread *td)
2962 {
2963 struct vnode *vp, *mvp, *rootvp = NULL;
2964 struct vattr vattr;
2965 int busy = 0, error;
2966
2967 CTR4(KTR_VFS, "%s: mp %p with rootrefs %d and flags %d", __func__, mp,
2968 rootrefs, flags);
2969 if (rootrefs > 0) {
2970 KASSERT((flags & (SKIPSYSTEM | WRITECLOSE)) == 0,
2971 ("vflush: bad args"));
2972 /*
2973 * Get the filesystem root vnode. We can vput() it
2974 * immediately, since with rootrefs > 0, it won't go away.
2975 */
2976 if ((error = VFS_ROOT(mp, LK_EXCLUSIVE, &rootvp)) != 0) {
2977 CTR2(KTR_VFS, "%s: vfs_root lookup failed with %d",
2978 __func__, error);
2979 return (error);
2980 }
2981 vput(rootvp);
2982 }
2983 loop:
2984 MNT_VNODE_FOREACH_ALL(vp, mp, mvp) {
2985 vholdl(vp);
2986 error = vn_lock(vp, LK_INTERLOCK | LK_EXCLUSIVE);
2987 if (error) {
2988 vdrop(vp);
2989 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
2990 goto loop;
2991 }
2992 /*
2993 * Skip over a vnodes marked VV_SYSTEM.
2994 */
2995 if ((flags & SKIPSYSTEM) && (vp->v_vflag & VV_SYSTEM)) {
2996 VOP_UNLOCK(vp, 0);
2997 vdrop(vp);
2998 continue;
2999 }
3000 /*
3001 * If WRITECLOSE is set, flush out unlinked but still open
3002 * files (even if open only for reading) and regular file
3003 * vnodes open for writing.
3004 */
3005 if (flags & WRITECLOSE) {
3006 if (vp->v_object != NULL) {
3007 VM_OBJECT_WLOCK(vp->v_object);
3008 vm_object_page_clean(vp->v_object, 0, 0, 0);
3009 VM_OBJECT_WUNLOCK(vp->v_object);
3010 }
3011 error = VOP_FSYNC(vp, MNT_WAIT, td);
3012 if (error != 0) {
3013 VOP_UNLOCK(vp, 0);
3014 vdrop(vp);
3015 MNT_VNODE_FOREACH_ALL_ABORT(mp, mvp);
3016 return (error);
3017 }
3018 error = VOP_GETATTR(vp, &vattr, td->td_ucred);
3019 VI_LOCK(vp);
3020
3021 if ((vp->v_type == VNON ||
3022 (error == 0 && vattr.va_nlink > 0)) &&
3023 (vp->v_writecount == 0 || vp->v_type != VREG)) {
3024 VOP_UNLOCK(vp, 0);
3025 vdropl(vp);
3026 continue;
3027 }
3028 } else
3029 VI_LOCK(vp);
3030 /*
3031 * With v_usecount == 0, all we need to do is clear out the
3032 * vnode data structures and we are done.
3033 *
3034 * If FORCECLOSE is set, forcibly close the vnode.
3035 */
3036 if (vp->v_usecount == 0 || (flags & FORCECLOSE)) {
3037 vgonel(vp);
3038 } else {
3039 busy++;
3040 #ifdef DIAGNOSTIC
3041 if (busyprt)
3042 vn_printf(vp, "vflush: busy vnode ");
3043 #endif
3044 }
3045 VOP_UNLOCK(vp, 0);
3046 vdropl(vp);
3047 }
3048 if (rootrefs > 0 && (flags & FORCECLOSE) == 0) {
3049 /*
3050 * If just the root vnode is busy, and if its refcount
3051 * is equal to `rootrefs', then go ahead and kill it.
3052 */
3053 VI_LOCK(rootvp);
3054 KASSERT(busy > 0, ("vflush: not busy"));
3055 VNASSERT(rootvp->v_usecount >= rootrefs, rootvp,
3056 ("vflush: usecount %d < rootrefs %d",
3057 rootvp->v_usecount, rootrefs));
3058 if (busy == 1 && rootvp->v_usecount == rootrefs) {
3059 VOP_LOCK(rootvp, LK_EXCLUSIVE|LK_INTERLOCK);
3060 vgone(rootvp);
3061 VOP_UNLOCK(rootvp, 0);
3062 busy = 0;
3063 } else
3064 VI_UNLOCK(rootvp);
3065 }
3066 if (busy) {
3067 CTR2(KTR_VFS, "%s: failing as %d vnodes are busy", __func__,
3068 busy);
3069 return (EBUSY);
3070 }
3071 for (; rootrefs > 0; rootrefs--)
3072 vrele(rootvp);
3073 return (0);
3074 }
3075
3076 /*
3077 * Recycle an unused vnode to the front of the free list.
3078 */
3079 int
3080 vrecycle(struct vnode *vp)
3081 {
3082 int recycled;
3083
3084 ASSERT_VOP_ELOCKED(vp, "vrecycle");
3085 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3086 recycled = 0;
3087 VI_LOCK(vp);
3088 if (vp->v_usecount == 0) {
3089 recycled = 1;
3090 vgonel(vp);
3091 }
3092 VI_UNLOCK(vp);
3093 return (recycled);
3094 }
3095
3096 /*
3097 * Eliminate all activity associated with a vnode
3098 * in preparation for reuse.
3099 */
3100 void
3101 vgone(struct vnode *vp)
3102 {
3103 VI_LOCK(vp);
3104 vgonel(vp);
3105 VI_UNLOCK(vp);
3106 }
3107
3108 static void
3109 notify_lowervp_vfs_dummy(struct mount *mp __unused,
3110 struct vnode *lowervp __unused)
3111 {
3112 }
3113
3114 /*
3115 * Notify upper mounts about reclaimed or unlinked vnode.
3116 */
3117 void
3118 vfs_notify_upper(struct vnode *vp, int event)
3119 {
3120 static struct vfsops vgonel_vfsops = {
3121 .vfs_reclaim_lowervp = notify_lowervp_vfs_dummy,
3122 .vfs_unlink_lowervp = notify_lowervp_vfs_dummy,
3123 };
3124 struct mount *mp, *ump, *mmp;
3125
3126 mp = vp->v_mount;
3127 if (mp == NULL)
3128 return;
3129
3130 MNT_ILOCK(mp);
3131 if (TAILQ_EMPTY(&mp->mnt_uppers))
3132 goto unlock;
3133 MNT_IUNLOCK(mp);
3134 mmp = malloc(sizeof(struct mount), M_TEMP, M_WAITOK | M_ZERO);
3135 mmp->mnt_op = &vgonel_vfsops;
3136 mmp->mnt_kern_flag |= MNTK_MARKER;
3137 MNT_ILOCK(mp);
3138 mp->mnt_kern_flag |= MNTK_VGONE_UPPER;
3139 for (ump = TAILQ_FIRST(&mp->mnt_uppers); ump != NULL;) {
3140 if ((ump->mnt_kern_flag & MNTK_MARKER) != 0) {
3141 ump = TAILQ_NEXT(ump, mnt_upper_link);
3142 continue;
3143 }
3144 TAILQ_INSERT_AFTER(&mp->mnt_uppers, ump, mmp, mnt_upper_link);
3145 MNT_IUNLOCK(mp);
3146 switch (event) {
3147 case VFS_NOTIFY_UPPER_RECLAIM:
3148 VFS_RECLAIM_LOWERVP(ump, vp);
3149 break;
3150 case VFS_NOTIFY_UPPER_UNLINK:
3151 VFS_UNLINK_LOWERVP(ump, vp);
3152 break;
3153 default:
3154 KASSERT(0, ("invalid event %d", event));
3155 break;
3156 }
3157 MNT_ILOCK(mp);
3158 ump = TAILQ_NEXT(mmp, mnt_upper_link);
3159 TAILQ_REMOVE(&mp->mnt_uppers, mmp, mnt_upper_link);
3160 }
3161 free(mmp, M_TEMP);
3162 mp->mnt_kern_flag &= ~MNTK_VGONE_UPPER;
3163 if ((mp->mnt_kern_flag & MNTK_VGONE_WAITER) != 0) {
3164 mp->mnt_kern_flag &= ~MNTK_VGONE_WAITER;
3165 wakeup(&mp->mnt_uppers);
3166 }
3167 unlock:
3168 MNT_IUNLOCK(mp);
3169 }
3170
3171 /*
3172 * vgone, with the vp interlock held.
3173 */
3174 static void
3175 vgonel(struct vnode *vp)
3176 {
3177 struct thread *td;
3178 int oweinact;
3179 int active;
3180 struct mount *mp;
3181
3182 ASSERT_VOP_ELOCKED(vp, "vgonel");
3183 ASSERT_VI_LOCKED(vp, "vgonel");
3184 VNASSERT(vp->v_holdcnt, vp,
3185 ("vgonel: vp %p has no reference.", vp));
3186 CTR2(KTR_VFS, "%s: vp %p", __func__, vp);
3187 td = curthread;
3188
3189 /*
3190 * Don't vgonel if we're already doomed.
3191 */
3192 if (vp->v_iflag & VI_DOOMED)
3193 return;
3194 vp->v_iflag |= VI_DOOMED;
3195
3196 /*
3197 * Check to see if the vnode is in use. If so, we have to call
3198 * VOP_CLOSE() and VOP_INACTIVE().
3199 */
3200 active = vp->v_usecount;
3201 oweinact = (vp->v_iflag & VI_OWEINACT);
3202 VI_UNLOCK(vp);
3203 vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
3204
3205 /*
3206 * If purging an active vnode, it must be closed and
3207 * deactivated before being reclaimed.
3208 */
3209 if (active)
3210 VOP_CLOSE(vp, FNONBLOCK, NOCRED, td);
3211 if (oweinact || active) {
3212 VI_LOCK(vp);
3213 if ((vp->v_iflag & VI_DOINGINACT) == 0)
3214 vinactive(vp, td);
3215 VI_UNLOCK(vp);
3216 }
3217 if (vp->v_type == VSOCK)
3218 vfs_unp_reclaim(vp);
3219
3220 /*
3221 * Clean out any buffers associated with the vnode.
3222 * If the flush fails, just toss the buffers.
3223 */
3224 mp = NULL;
3225 if (!TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd))
3226 (void) vn_start_secondary_write(vp, &mp, V_WAIT);
3227 if (vinvalbuf(vp, V_SAVE, 0, 0) != 0) {
3228 while (vinvalbuf(vp, 0, 0, 0) != 0)
3229 ;
3230 }
3231
3232 BO_LOCK(&vp->v_bufobj);
3233 KASSERT(TAILQ_EMPTY(&vp->v_bufobj.bo_dirty.bv_hd) &&
3234 vp->v_bufobj.bo_dirty.bv_cnt == 0 &&
3235 TAILQ_EMPTY(&vp->v_bufobj.bo_clean.bv_hd) &&
3236 vp->v_bufobj.bo_clean.bv_cnt == 0,
3237 ("vp %p bufobj not invalidated", vp));
3238
3239 /*
3240 * For VMIO bufobj, BO_DEAD is set in vm_object_terminate()
3241 * after the object's page queue is flushed.
3242 */
3243 if (vp->v_bufobj.bo_object == NULL)
3244 vp->v_bufobj.bo_flag |= BO_DEAD;
3245 BO_UNLOCK(&vp->v_bufobj);
3246
3247 /*
3248 * Reclaim the vnode.
3249 */
3250 if (VOP_RECLAIM(vp, td))
3251 panic("vgone: cannot reclaim");
3252 if (mp != NULL)
3253 vn_finished_secondary_write(mp);
3254 VNASSERT(vp->v_object == NULL, vp,
3255 ("vop_reclaim left v_object vp=%p, tag=%s", vp, vp->v_tag));
3256 /*
3257 * Clear the advisory locks and wake up waiting threads.
3258 */
3259 (void)VOP_ADVLOCKPURGE(vp);
3260 vp->v_lockf = NULL;
3261 /*
3262 * Delete from old mount point vnode list.
3263 */
3264 delmntque(vp);
3265 cache_purge(vp);
3266 /*
3267 * Done with purge, reset to the standard lock and invalidate
3268 * the vnode.
3269 */
3270 VI_LOCK(vp);
3271 vp->v_vnlock = &vp->v_lock;
3272 vp->v_op = &dead_vnodeops;
3273 vp->v_tag = "none";
3274 vp->v_type = VBAD;
3275 }
3276
3277 /*
3278 * Calculate the total number of references to a special device.
3279 */
3280 int
3281 vcount(struct vnode *vp)
3282 {
3283 int count;
3284
3285 dev_lock();
3286 count = vp->v_rdev->si_usecount;
3287 dev_unlock();
3288 return (count);
3289 }
3290
3291 /*
3292 * Same as above, but using the struct cdev *as argument
3293 */
3294 int
3295 count_dev(struct cdev *dev)
3296 {
3297 int count;
3298
3299 dev_lock();
3300 count = dev->si_usecount;
3301 dev_unlock();
3302 return(count);
3303 }
3304
3305 /*
3306 * Print out a description of a vnode.
3307 */
3308 static char *typename[] =
3309 {"VNON", "VREG", "VDIR", "VBLK", "VCHR", "VLNK", "VSOCK", "VFIFO", "VBAD",
3310 "VMARKER"};
3311
3312 void
3313 vn_printf(struct vnode *vp, const char *fmt, ...)
3314 {
3315 va_list ap;
3316 char buf[256], buf2[16];
3317 u_long flags;
3318
3319 va_start(ap, fmt);
3320 vprintf(fmt, ap);
3321 va_end(ap);
3322 printf("%p: ", (void *)vp);
3323 printf("tag %s, type %s\n", vp->v_tag, typename[vp->v_type]);
3324 printf(" usecount %d, writecount %d, refcount %d",
3325 vp->v_usecount, vp->v_writecount, vp->v_holdcnt);
3326 switch (vp->v_type) {
3327 case VDIR:
3328 printf(" mountedhere %p\n", vp->v_mountedhere);
3329 break;
3330 case VCHR:
3331 printf(" rdev %p\n", vp->v_rdev);
3332 break;
3333 case VSOCK:
3334 printf(" socket %p\n", vp->v_socket);
3335 break;
3336 case VFIFO:
3337 printf(" fifoinfo %p\n", vp->v_fifoinfo);
3338 break;
3339 default:
3340 printf("\n");
3341 break;
3342 }
3343 buf[0] = '\0';
3344 buf[1] = '\0';
3345 if (vp->v_vflag & VV_ROOT)
3346 strlcat(buf, "|VV_ROOT", sizeof(buf));
3347 if (vp->v_vflag & VV_ISTTY)
3348 strlcat(buf, "|VV_ISTTY", sizeof(buf));
3349 if (vp->v_vflag & VV_NOSYNC)
3350 strlcat(buf, "|VV_NOSYNC", sizeof(buf));
3351 if (vp->v_vflag & VV_ETERNALDEV)
3352 strlcat(buf, "|VV_ETERNALDEV", sizeof(buf));
3353 if (vp->v_vflag & VV_CACHEDLABEL)
3354 strlcat(buf, "|VV_CACHEDLABEL", sizeof(buf));
3355 if (vp->v_vflag & VV_TEXT)
3356 strlcat(buf, "|VV_TEXT", sizeof(buf));
3357 if (vp->v_vflag & VV_COPYONWRITE)
3358 strlcat(buf, "|VV_COPYONWRITE", sizeof(buf));
3359 if (vp->v_vflag & VV_SYSTEM)
3360 strlcat(buf, "|VV_SYSTEM", sizeof(buf));
3361 if (vp->v_vflag & VV_PROCDEP)
3362 strlcat(buf, "|VV_PROCDEP", sizeof(buf));
3363 if (vp->v_vflag & VV_NOKNOTE)
3364 strlcat(buf, "|VV_NOKNOTE", sizeof(buf));
3365 if (vp->v_vflag & VV_DELETED)
3366 strlcat(buf, "|VV_DELETED", sizeof(buf));
3367 if (vp->v_vflag & VV_MD)
3368 strlcat(buf, "|VV_MD", sizeof(buf));
3369 if (vp->v_vflag & VV_FORCEINSMQ)
3370 strlcat(buf, "|VV_FORCEINSMQ", sizeof(buf));
3371 flags = vp->v_vflag & ~(VV_ROOT | VV_ISTTY | VV_NOSYNC | VV_ETERNALDEV |
3372 VV_CACHEDLABEL | VV_TEXT | VV_COPYONWRITE | VV_SYSTEM | VV_PROCDEP |
3373 VV_NOKNOTE | VV_DELETED | VV_MD | VV_FORCEINSMQ);
3374 if (flags != 0) {
3375 snprintf(buf2, sizeof(buf2), "|VV(0x%lx)", flags);
3376 strlcat(buf, buf2, sizeof(buf));
3377 }
3378 if (vp->v_iflag & VI_MOUNT)
3379 strlcat(buf, "|VI_MOUNT", sizeof(buf));
3380 if (vp->v_iflag & VI_DOOMED)
3381 strlcat(buf, "|VI_DOOMED", sizeof(buf));
3382 if (vp->v_iflag & VI_FREE)
3383 strlcat(buf, "|VI_FREE", sizeof(buf));
3384 if (vp->v_iflag & VI_ACTIVE)
3385 strlcat(buf, "|VI_ACTIVE", sizeof(buf));
3386 if (vp->v_iflag & VI_DOINGINACT)
3387 strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
3388 if (vp->v_iflag & VI_OWEINACT)
3389 strlcat(buf, "|VI_OWEINACT", sizeof(buf));
3390 flags = vp->v_iflag & ~(VI_MOUNT | VI_DOOMED | VI_FREE |
3391 VI_ACTIVE | VI_DOINGINACT | VI_OWEINACT);
3392 if (flags != 0) {
3393 snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
3394 strlcat(buf, buf2, sizeof(buf));
3395 }
3396 printf(" flags (%s)\n", buf + 1);
3397 if (mtx_owned(VI_MTX(vp)))
3398 printf(" VI_LOCKed");
3399 if (vp->v_object != NULL)
3400 printf(" v_object %p ref %d pages %d "
3401 "cleanbuf %d dirtybuf %d\n",
3402 vp->v_object, vp->v_object->ref_count,
3403 vp->v_object->resident_page_count,
3404 vp->v_bufobj.bo_clean.bv_cnt,
3405 vp->v_bufobj.bo_dirty.bv_cnt);
3406 printf(" ");
3407 lockmgr_printinfo(vp->v_vnlock);
3408 if (vp->v_data != NULL)
3409 VOP_PRINT(vp);
3410 }
3411
3412 #ifdef DDB
3413 /*
3414 * List all of the locked vnodes in the system.
3415 * Called when debugging the kernel.
3416 */
3417 DB_SHOW_COMMAND(lockedvnods, lockedvnodes)
3418 {
3419 struct mount *mp;
3420 struct vnode *vp;
3421
3422 /*
3423 * Note: because this is DDB, we can't obey the locking semantics
3424 * for these structures, which means we could catch an inconsistent
3425 * state and dereference a nasty pointer. Not much to be done
3426 * about that.
3427 */
3428 db_printf("Locked vnodes\n");
3429 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3430 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3431 if (vp->v_type != VMARKER && VOP_ISLOCKED(vp))
3432 vn_printf(vp, "vnode ");
3433 }
3434 }
3435 }
3436
3437 /*
3438 * Show details about the given vnode.
3439 */
3440 DB_SHOW_COMMAND(vnode, db_show_vnode)
3441 {
3442 struct vnode *vp;
3443
3444 if (!have_addr)
3445 return;
3446 vp = (struct vnode *)addr;
3447 vn_printf(vp, "vnode ");
3448 }
3449
3450 /*
3451 * Show details about the given mount point.
3452 */
3453 DB_SHOW_COMMAND(mount, db_show_mount)
3454 {
3455 struct mount *mp;
3456 struct vfsopt *opt;
3457 struct statfs *sp;
3458 struct vnode *vp;
3459 char buf[512];
3460 uint64_t mflags;
3461 u_int flags;
3462
3463 if (!have_addr) {
3464 /* No address given, print short info about all mount points. */
3465 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3466 db_printf("%p %s on %s (%s)\n", mp,
3467 mp->mnt_stat.f_mntfromname,
3468 mp->mnt_stat.f_mntonname,
3469 mp->mnt_stat.f_fstypename);
3470 if (db_pager_quit)
3471 break;
3472 }
3473 db_printf("\nMore info: show mount <addr>\n");
3474 return;
3475 }
3476
3477 mp = (struct mount *)addr;
3478 db_printf("%p %s on %s (%s)\n", mp, mp->mnt_stat.f_mntfromname,
3479 mp->mnt_stat.f_mntonname, mp->mnt_stat.f_fstypename);
3480
3481 buf[0] = '\0';
3482 mflags = mp->mnt_flag;
3483 #define MNT_FLAG(flag) do { \
3484 if (mflags & (flag)) { \
3485 if (buf[0] != '\0') \
3486 strlcat(buf, ", ", sizeof(buf)); \
3487 strlcat(buf, (#flag) + 4, sizeof(buf)); \
3488 mflags &= ~(flag); \
3489 } \
3490 } while (0)
3491 MNT_FLAG(MNT_RDONLY);
3492 MNT_FLAG(MNT_SYNCHRONOUS);
3493 MNT_FLAG(MNT_NOEXEC);
3494 MNT_FLAG(MNT_NOSUID);
3495 MNT_FLAG(MNT_NFS4ACLS);
3496 MNT_FLAG(MNT_UNION);
3497 MNT_FLAG(MNT_ASYNC);
3498 MNT_FLAG(MNT_SUIDDIR);
3499 MNT_FLAG(MNT_SOFTDEP);
3500 MNT_FLAG(MNT_NOSYMFOLLOW);
3501 MNT_FLAG(MNT_GJOURNAL);
3502 MNT_FLAG(MNT_MULTILABEL);
3503 MNT_FLAG(MNT_ACLS);
3504 MNT_FLAG(MNT_NOATIME);
3505 MNT_FLAG(MNT_NOCLUSTERR);
3506 MNT_FLAG(MNT_NOCLUSTERW);
3507 MNT_FLAG(MNT_SUJ);
3508 MNT_FLAG(MNT_EXRDONLY);
3509 MNT_FLAG(MNT_EXPORTED);
3510 MNT_FLAG(MNT_DEFEXPORTED);
3511 MNT_FLAG(MNT_EXPORTANON);
3512 MNT_FLAG(MNT_EXKERB);
3513 MNT_FLAG(MNT_EXPUBLIC);
3514 MNT_FLAG(MNT_LOCAL);
3515 MNT_FLAG(MNT_QUOTA);
3516 MNT_FLAG(MNT_ROOTFS);
3517 MNT_FLAG(MNT_USER);
3518 MNT_FLAG(MNT_IGNORE);
3519 MNT_FLAG(MNT_UPDATE);
3520 MNT_FLAG(MNT_DELEXPORT);
3521 MNT_FLAG(MNT_RELOAD);
3522 MNT_FLAG(MNT_FORCE);
3523 MNT_FLAG(MNT_SNAPSHOT);
3524 MNT_FLAG(MNT_BYFSID);
3525 #undef MNT_FLAG
3526 if (mflags != 0) {
3527 if (buf[0] != '\0')
3528 strlcat(buf, ", ", sizeof(buf));
3529 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3530 "0x%016jx", mflags);
3531 }
3532 db_printf(" mnt_flag = %s\n", buf);
3533
3534 buf[0] = '\0';
3535 flags = mp->mnt_kern_flag;
3536 #define MNT_KERN_FLAG(flag) do { \
3537 if (flags & (flag)) { \
3538 if (buf[0] != '\0') \
3539 strlcat(buf, ", ", sizeof(buf)); \
3540 strlcat(buf, (#flag) + 5, sizeof(buf)); \
3541 flags &= ~(flag); \
3542 } \
3543 } while (0)
3544 MNT_KERN_FLAG(MNTK_UNMOUNTF);
3545 MNT_KERN_FLAG(MNTK_ASYNC);
3546 MNT_KERN_FLAG(MNTK_SOFTDEP);
3547 MNT_KERN_FLAG(MNTK_NOINSMNTQ);
3548 MNT_KERN_FLAG(MNTK_DRAINING);
3549 MNT_KERN_FLAG(MNTK_REFEXPIRE);
3550 MNT_KERN_FLAG(MNTK_EXTENDED_SHARED);
3551 MNT_KERN_FLAG(MNTK_SHARED_WRITES);
3552 MNT_KERN_FLAG(MNTK_NO_IOPF);
3553 MNT_KERN_FLAG(MNTK_VGONE_UPPER);
3554 MNT_KERN_FLAG(MNTK_VGONE_WAITER);
3555 MNT_KERN_FLAG(MNTK_LOOKUP_EXCL_DOTDOT);
3556 MNT_KERN_FLAG(MNTK_MARKER);
3557 MNT_KERN_FLAG(MNTK_USES_BCACHE);
3558 MNT_KERN_FLAG(MNTK_NOASYNC);
3559 MNT_KERN_FLAG(MNTK_UNMOUNT);
3560 MNT_KERN_FLAG(MNTK_MWAIT);
3561 MNT_KERN_FLAG(MNTK_SUSPEND);
3562 MNT_KERN_FLAG(MNTK_SUSPEND2);
3563 MNT_KERN_FLAG(MNTK_SUSPENDED);
3564 MNT_KERN_FLAG(MNTK_LOOKUP_SHARED);
3565 MNT_KERN_FLAG(MNTK_NOKNOTE);
3566 #undef MNT_KERN_FLAG
3567 if (flags != 0) {
3568 if (buf[0] != '\0')
3569 strlcat(buf, ", ", sizeof(buf));
3570 snprintf(buf + strlen(buf), sizeof(buf) - strlen(buf),
3571 "0x%08x", flags);
3572 }
3573 db_printf(" mnt_kern_flag = %s\n", buf);
3574
3575 db_printf(" mnt_opt = ");
3576 opt = TAILQ_FIRST(mp->mnt_opt);
3577 if (opt != NULL) {
3578 db_printf("%s", opt->name);
3579 opt = TAILQ_NEXT(opt, link);
3580 while (opt != NULL) {
3581 db_printf(", %s", opt->name);
3582 opt = TAILQ_NEXT(opt, link);
3583 }
3584 }
3585 db_printf("\n");
3586
3587 sp = &mp->mnt_stat;
3588 db_printf(" mnt_stat = { version=%u type=%u flags=0x%016jx "
3589 "bsize=%ju iosize=%ju blocks=%ju bfree=%ju bavail=%jd files=%ju "
3590 "ffree=%jd syncwrites=%ju asyncwrites=%ju syncreads=%ju "
3591 "asyncreads=%ju namemax=%u owner=%u fsid=[%d, %d] }\n",
3592 (u_int)sp->f_version, (u_int)sp->f_type, (uintmax_t)sp->f_flags,
3593 (uintmax_t)sp->f_bsize, (uintmax_t)sp->f_iosize,
3594 (uintmax_t)sp->f_blocks, (uintmax_t)sp->f_bfree,
3595 (intmax_t)sp->f_bavail, (uintmax_t)sp->f_files,
3596 (intmax_t)sp->f_ffree, (uintmax_t)sp->f_syncwrites,
3597 (uintmax_t)sp->f_asyncwrites, (uintmax_t)sp->f_syncreads,
3598 (uintmax_t)sp->f_asyncreads, (u_int)sp->f_namemax,
3599 (u_int)sp->f_owner, (int)sp->f_fsid.val[0], (int)sp->f_fsid.val[1]);
3600
3601 db_printf(" mnt_cred = { uid=%u ruid=%u",
3602 (u_int)mp->mnt_cred->cr_uid, (u_int)mp->mnt_cred->cr_ruid);
3603 if (jailed(mp->mnt_cred))
3604 db_printf(", jail=%d", mp->mnt_cred->cr_prison->pr_id);
3605 db_printf(" }\n");
3606 db_printf(" mnt_ref = %d\n", mp->mnt_ref);
3607 db_printf(" mnt_gen = %d\n", mp->mnt_gen);
3608 db_printf(" mnt_nvnodelistsize = %d\n", mp->mnt_nvnodelistsize);
3609 db_printf(" mnt_activevnodelistsize = %d\n",
3610 mp->mnt_activevnodelistsize);
3611 db_printf(" mnt_writeopcount = %d\n", mp->mnt_writeopcount);
3612 db_printf(" mnt_maxsymlinklen = %d\n", mp->mnt_maxsymlinklen);
3613 db_printf(" mnt_iosize_max = %d\n", mp->mnt_iosize_max);
3614 db_printf(" mnt_hashseed = %u\n", mp->mnt_hashseed);
3615 db_printf(" mnt_lockref = %d\n", mp->mnt_lockref);
3616 db_printf(" mnt_secondary_writes = %d\n", mp->mnt_secondary_writes);
3617 db_printf(" mnt_secondary_accwrites = %d\n",
3618 mp->mnt_secondary_accwrites);
3619 db_printf(" mnt_gjprovider = %s\n",
3620 mp->mnt_gjprovider != NULL ? mp->mnt_gjprovider : "NULL");
3621
3622 db_printf("\n\nList of active vnodes\n");
3623 TAILQ_FOREACH(vp, &mp->mnt_activevnodelist, v_actfreelist) {
3624 if (vp->v_type != VMARKER) {
3625 vn_printf(vp, "vnode ");
3626 if (db_pager_quit)
3627 break;
3628 }
3629 }
3630 db_printf("\n\nList of inactive vnodes\n");
3631 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3632 if (vp->v_type != VMARKER && (vp->v_iflag & VI_ACTIVE) == 0) {
3633 vn_printf(vp, "vnode ");
3634 if (db_pager_quit)
3635 break;
3636 }
3637 }
3638 }
3639 #endif /* DDB */
3640
3641 /*
3642 * Fill in a struct xvfsconf based on a struct vfsconf.
3643 */
3644 static int
3645 vfsconf2x(struct sysctl_req *req, struct vfsconf *vfsp)
3646 {
3647 struct xvfsconf xvfsp;
3648
3649 bzero(&xvfsp, sizeof(xvfsp));
3650 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3651 xvfsp.vfc_typenum = vfsp->vfc_typenum;
3652 xvfsp.vfc_refcount = vfsp->vfc_refcount;
3653 xvfsp.vfc_flags = vfsp->vfc_flags;
3654 /*
3655 * These are unused in userland, we keep them
3656 * to not break binary compatibility.
3657 */
3658 xvfsp.vfc_vfsops = NULL;
3659 xvfsp.vfc_next = NULL;
3660 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3661 }
3662
3663 #ifdef COMPAT_FREEBSD32
3664 struct xvfsconf32 {
3665 uint32_t vfc_vfsops;
3666 char vfc_name[MFSNAMELEN];
3667 int32_t vfc_typenum;
3668 int32_t vfc_refcount;
3669 int32_t vfc_flags;
3670 uint32_t vfc_next;
3671 };
3672
3673 static int
3674 vfsconf2x32(struct sysctl_req *req, struct vfsconf *vfsp)
3675 {
3676 struct xvfsconf32 xvfsp;
3677
3678 bzero(&xvfsp, sizeof(xvfsp));
3679 strcpy(xvfsp.vfc_name, vfsp->vfc_name);
3680 xvfsp.vfc_typenum = vfsp->vfc_typenum;
3681 xvfsp.vfc_refcount = vfsp->vfc_refcount;
3682 xvfsp.vfc_flags = vfsp->vfc_flags;
3683 return (SYSCTL_OUT(req, &xvfsp, sizeof(xvfsp)));
3684 }
3685 #endif
3686
3687 /*
3688 * Top level filesystem related information gathering.
3689 */
3690 static int
3691 sysctl_vfs_conflist(SYSCTL_HANDLER_ARGS)
3692 {
3693 struct vfsconf *vfsp;
3694 int error;
3695
3696 error = 0;
3697 vfsconf_slock();
3698 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3699 #ifdef COMPAT_FREEBSD32
3700 if (req->flags & SCTL_MASK32)
3701 error = vfsconf2x32(req, vfsp);
3702 else
3703 #endif
3704 error = vfsconf2x(req, vfsp);
3705 if (error)
3706 break;
3707 }
3708 vfsconf_sunlock();
3709 return (error);
3710 }
3711
3712 SYSCTL_PROC(_vfs, OID_AUTO, conflist, CTLTYPE_OPAQUE | CTLFLAG_RD |
3713 CTLFLAG_MPSAFE, NULL, 0, sysctl_vfs_conflist,
3714 "S,xvfsconf", "List of all configured filesystems");
3715
3716 #ifndef BURN_BRIDGES
3717 static int sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS);
3718
3719 static int
3720 vfs_sysctl(SYSCTL_HANDLER_ARGS)
3721 {
3722 int *name = (int *)arg1 - 1; /* XXX */
3723 u_int namelen = arg2 + 1; /* XXX */
3724 struct vfsconf *vfsp;
3725
3726 log(LOG_WARNING, "userland calling deprecated sysctl, "
3727 "please rebuild world\n");
3728
3729 #if 1 || defined(COMPAT_PRELITE2)
3730 /* Resolve ambiguity between VFS_VFSCONF and VFS_GENERIC. */
3731 if (namelen == 1)
3732 return (sysctl_ovfs_conf(oidp, arg1, arg2, req));
3733 #endif
3734
3735 switch (name[1]) {
3736 case VFS_MAXTYPENUM:
3737 if (namelen != 2)
3738 return (ENOTDIR);
3739 return (SYSCTL_OUT(req, &maxvfsconf, sizeof(int)));
3740 case VFS_CONF:
3741 if (namelen != 3)
3742 return (ENOTDIR); /* overloaded */
3743 vfsconf_slock();
3744 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3745 if (vfsp->vfc_typenum == name[2])
3746 break;
3747 }
3748 vfsconf_sunlock();
3749 if (vfsp == NULL)
3750 return (EOPNOTSUPP);
3751 #ifdef COMPAT_FREEBSD32
3752 if (req->flags & SCTL_MASK32)
3753 return (vfsconf2x32(req, vfsp));
3754 else
3755 #endif
3756 return (vfsconf2x(req, vfsp));
3757 }
3758 return (EOPNOTSUPP);
3759 }
3760
3761 static SYSCTL_NODE(_vfs, VFS_GENERIC, generic, CTLFLAG_RD | CTLFLAG_SKIP |
3762 CTLFLAG_MPSAFE, vfs_sysctl,
3763 "Generic filesystem");
3764
3765 #if 1 || defined(COMPAT_PRELITE2)
3766
3767 static int
3768 sysctl_ovfs_conf(SYSCTL_HANDLER_ARGS)
3769 {
3770 int error;
3771 struct vfsconf *vfsp;
3772 struct ovfsconf ovfs;
3773
3774 vfsconf_slock();
3775 TAILQ_FOREACH(vfsp, &vfsconf, vfc_list) {
3776 bzero(&ovfs, sizeof(ovfs));
3777 ovfs.vfc_vfsops = vfsp->vfc_vfsops; /* XXX used as flag */
3778 strcpy(ovfs.vfc_name, vfsp->vfc_name);
3779 ovfs.vfc_index = vfsp->vfc_typenum;
3780 ovfs.vfc_refcount = vfsp->vfc_refcount;
3781 ovfs.vfc_flags = vfsp->vfc_flags;
3782 error = SYSCTL_OUT(req, &ovfs, sizeof ovfs);
3783 if (error != 0) {
3784 vfsconf_sunlock();
3785 return (error);
3786 }
3787 }
3788 vfsconf_sunlock();
3789 return (0);
3790 }
3791
3792 #endif /* 1 || COMPAT_PRELITE2 */
3793 #endif /* !BURN_BRIDGES */
3794
3795 #define KINFO_VNODESLOP 10
3796 #ifdef notyet
3797 /*
3798 * Dump vnode list (via sysctl).
3799 */
3800 /* ARGSUSED */
3801 static int
3802 sysctl_vnode(SYSCTL_HANDLER_ARGS)
3803 {
3804 struct xvnode *xvn;
3805 struct mount *mp;
3806 struct vnode *vp;
3807 int error, len, n;
3808
3809 /*
3810 * Stale numvnodes access is not fatal here.
3811 */
3812 req->lock = 0;
3813 len = (numvnodes + KINFO_VNODESLOP) * sizeof *xvn;
3814 if (!req->oldptr)
3815 /* Make an estimate */
3816 return (SYSCTL_OUT(req, 0, len));
3817
3818 error = sysctl_wire_old_buffer(req, 0);
3819 if (error != 0)
3820 return (error);
3821 xvn = malloc(len, M_TEMP, M_ZERO | M_WAITOK);
3822 n = 0;
3823 mtx_lock(&mountlist_mtx);
3824 TAILQ_FOREACH(mp, &mountlist, mnt_list) {
3825 if (vfs_busy(mp, MBF_NOWAIT | MBF_MNTLSTLOCK))
3826 continue;
3827 MNT_ILOCK(mp);
3828 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
3829 if (n == len)
3830 break;
3831 vref(vp);
3832 xvn[n].xv_size = sizeof *xvn;
3833 xvn[n].xv_vnode = vp;
3834 xvn[n].xv_id = 0; /* XXX compat */
3835 #define XV_COPY(field) xvn[n].xv_##field = vp->v_##field
3836 XV_COPY(usecount);
3837 XV_COPY(writecount);
3838 XV_COPY(holdcnt);
3839 XV_COPY(mount);
3840 XV_COPY(numoutput);
3841 XV_COPY(type);
3842 #undef XV_COPY
3843 xvn[n].xv_flag = vp->v_vflag;
3844
3845 switch (vp->v_type) {
3846 case VREG:
3847 case VDIR:
3848 case VLNK:
3849 break;
3850 case VBLK:
3851 case VCHR:
3852 if (vp->v_rdev == NULL) {
3853 vrele(vp);
3854 continue;
3855 }
3856 xvn[n].xv_dev = dev2udev(vp->v_rdev);
3857 break;
3858 case VSOCK:
3859 xvn[n].xv_socket = vp->v_socket;
3860 break;
3861 case VFIFO:
3862 xvn[n].xv_fifo = vp->v_fifoinfo;
3863 break;
3864 case VNON:
3865 case VBAD:
3866 default:
3867 /* shouldn't happen? */
3868 vrele(vp);
3869 continue;
3870 }
3871 vrele(vp);
3872 ++n;
3873 }
3874 MNT_IUNLOCK(mp);
3875 mtx_lock(&mountlist_mtx);
3876 vfs_unbusy(mp);
3877 if (n == len)
3878 break;
3879 }
3880 mtx_unlock(&mountlist_mtx);
3881
3882 error = SYSCTL_OUT(req, xvn, n * sizeof *xvn);
3883 free(xvn, M_TEMP);
3884 return (error);
3885 }
3886
3887 SYSCTL_PROC(_kern, KERN_VNODE, vnode, CTLTYPE_OPAQUE | CTLFLAG_RD |
3888 CTLFLAG_MPSAFE, 0, 0, sysctl_vnode, "S,xvnode",
3889 "");
3890 #endif
3891
3892 static void
3893 unmount_or_warn(struct mount *mp)
3894 {
3895 int error;
3896
3897 error = dounmount(mp, MNT_FORCE, curthread);
3898 if (error != 0) {
3899 printf("unmount of %s failed (", mp->mnt_stat.f_mntonname);
3900 if (error == EBUSY)
3901 printf("BUSY)\n");
3902 else
3903 printf("%d)\n", error);
3904 }
3905 }
3906
3907 /*
3908 * Unmount all filesystems. The list is traversed in reverse order
3909 * of mounting to avoid dependencies.
3910 */
3911 void
3912 vfs_unmountall(void)
3913 {
3914 struct mount *mp, *tmp;
3915
3916 CTR1(KTR_VFS, "%s: unmounting all filesystems", __func__);
3917
3918 /*
3919 * Since this only runs when rebooting, it is not interlocked.
3920 */
3921 TAILQ_FOREACH_REVERSE_SAFE(mp, &mountlist, mntlist, mnt_list, tmp) {
3922 vfs_ref(mp);
3923
3924 /*
3925 * Forcibly unmounting "/dev" before "/" would prevent clean
3926 * unmount of the latter.
3927 */
3928 if (mp == rootdevmp)
3929 continue;
3930
3931 unmount_or_warn(mp);
3932 }
3933
3934 if (rootdevmp != NULL)
3935 unmount_or_warn(rootdevmp);
3936 }
3937
3938 /*
3939 * perform msync on all vnodes under a mount point
3940 * the mount point must be locked.
3941 */
3942 void
3943 vfs_msync(struct mount *mp, int flags)
3944 {
3945 struct vnode *vp, *mvp;
3946 struct vm_object *obj;
3947
3948 CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
3949 MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
3950 obj = vp->v_object;
3951 if (obj != NULL && (obj->flags & OBJ_MIGHTBEDIRTY) != 0 &&
3952 (flags == MNT_WAIT || VOP_ISLOCKED(vp) == 0)) {
3953 if (!vget(vp,
3954 LK_EXCLUSIVE | LK_RETRY | LK_INTERLOCK,
3955 curthread)) {
3956 if (vp->v_vflag & VV_NOSYNC) { /* unlinked */
3957 vput(vp);
3958 continue;
3959 }
3960
3961 obj = vp->v_object;
3962 if (obj != NULL) {
3963 VM_OBJECT_WLOCK(obj);
3964 vm_object_page_clean(obj, 0, 0,
3965 flags == MNT_WAIT ?
3966 OBJPC_SYNC : OBJPC_NOSYNC);
3967 VM_OBJECT_WUNLOCK(obj);
3968 }
3969 vput(vp);
3970 }
3971 } else
3972 VI_UNLOCK(vp);
3973 }
3974 }
3975
3976 static void
3977 destroy_vpollinfo_free(struct vpollinfo *vi)
3978 {
3979
3980 knlist_destroy(&vi->vpi_selinfo.si_note);
3981 mtx_destroy(&vi->vpi_lock);
3982 uma_zfree(vnodepoll_zone, vi);
3983 }
3984
3985 static void
3986 destroy_vpollinfo(struct vpollinfo *vi)
3987 {
3988
3989 knlist_clear(&vi->vpi_selinfo.si_note, 1);
3990 seldrain(&vi->vpi_selinfo);
3991 destroy_vpollinfo_free(vi);
3992 }
3993
3994 /*
3995 * Initialize per-vnode helper structure to hold poll-related state.
3996 */
3997 void
3998 v_addpollinfo(struct vnode *vp)
3999 {
4000 struct vpollinfo *vi;
4001
4002 if (vp->v_pollinfo != NULL)
4003 return;
4004 vi = uma_zalloc(vnodepoll_zone, M_WAITOK | M_ZERO);
4005 mtx_init(&vi->vpi_lock, "vnode pollinfo", NULL, MTX_DEF);
4006 knlist_init(&vi->vpi_selinfo.si_note, vp, vfs_knllock,
4007 vfs_knlunlock, vfs_knl_assert_locked, vfs_knl_assert_unlocked);
4008 VI_LOCK(vp);
4009 if (vp->v_pollinfo != NULL) {
4010 VI_UNLOCK(vp);
4011 destroy_vpollinfo_free(vi);
4012 return;
4013 }
4014 vp->v_pollinfo = vi;
4015 VI_UNLOCK(vp);
4016 }
4017
4018 /*
4019 * Record a process's interest in events which might happen to
4020 * a vnode. Because poll uses the historic select-style interface
4021 * internally, this routine serves as both the ``check for any
4022 * pending events'' and the ``record my interest in future events''
4023 * functions. (These are done together, while the lock is held,
4024 * to avoid race conditions.)
4025 */
4026 int
4027 vn_pollrecord(struct vnode *vp, struct thread *td, int events)
4028 {
4029
4030 v_addpollinfo(vp);
4031 mtx_lock(&vp->v_pollinfo->vpi_lock);
4032 if (vp->v_pollinfo->vpi_revents & events) {
4033 /*
4034 * This leaves events we are not interested
4035 * in available for the other process which
4036 * which presumably had requested them
4037 * (otherwise they would never have been
4038 * recorded).
4039 */
4040 events &= vp->v_pollinfo->vpi_revents;
4041 vp->v_pollinfo->vpi_revents &= ~events;
4042
4043 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4044 return (events);
4045 }
4046 vp->v_pollinfo->vpi_events |= events;
4047 selrecord(td, &vp->v_pollinfo->vpi_selinfo);
4048 mtx_unlock(&vp->v_pollinfo->vpi_lock);
4049 return (0);
4050 }
4051
4052 /*
4053 * Routine to create and manage a filesystem syncer vnode.
4054 */
4055 #define sync_close ((int (*)(struct vop_close_args *))nullop)
4056 static int sync_fsync(struct vop_fsync_args *);
4057 static int sync_inactive(struct vop_inactive_args *);
4058 static int sync_reclaim(struct vop_reclaim_args *);
4059
4060 static struct vop_vector sync_vnodeops = {
4061 .vop_bypass = VOP_EOPNOTSUPP,
4062 .vop_close = sync_close, /* close */
4063 .vop_fsync = sync_fsync, /* fsync */
4064 .vop_inactive = sync_inactive, /* inactive */
4065 .vop_reclaim = sync_reclaim, /* reclaim */
4066 .vop_lock1 = vop_stdlock, /* lock */
4067 .vop_unlock = vop_stdunlock, /* unlock */
4068 .vop_islocked = vop_stdislocked, /* islocked */
4069 };
4070
4071 /*
4072 * Create a new filesystem syncer vnode for the specified mount point.
4073 */
4074 void
4075 vfs_allocate_syncvnode(struct mount *mp)
4076 {
4077 struct vnode *vp;
4078 struct bufobj *bo;
4079 static long start, incr, next;
4080 int error;
4081
4082 /* Allocate a new vnode */
4083 error = getnewvnode("syncer", mp, &sync_vnodeops, &vp);
4084 if (error != 0)
4085 panic("vfs_allocate_syncvnode: getnewvnode() failed");
4086 vp->v_type = VNON;
4087 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4088 vp->v_vflag |= VV_FORCEINSMQ;
4089 error = insmntque(vp, mp);
4090 if (error != 0)
4091 panic("vfs_allocate_syncvnode: insmntque() failed");
4092 vp->v_vflag &= ~VV_FORCEINSMQ;
4093 VOP_UNLOCK(vp, 0);
4094 /*
4095 * Place the vnode onto the syncer worklist. We attempt to
4096 * scatter them about on the list so that they will go off
4097 * at evenly distributed times even if all the filesystems
4098 * are mounted at once.
4099 */
4100 next += incr;
4101 if (next == 0 || next > syncer_maxdelay) {
4102 start /= 2;
4103 incr /= 2;
4104 if (start == 0) {
4105 start = syncer_maxdelay / 2;
4106 incr = syncer_maxdelay;
4107 }
4108 next = start;
4109 }
4110 bo = &vp->v_bufobj;
4111 BO_LOCK(bo);
4112 vn_syncer_add_to_worklist(bo, syncdelay > 0 ? next % syncdelay : 0);
4113 /* XXX - vn_syncer_add_to_worklist() also grabs and drops sync_mtx. */
4114 mtx_lock(&sync_mtx);
4115 sync_vnode_count++;
4116 if (mp->mnt_syncer == NULL) {
4117 mp->mnt_syncer = vp;
4118 vp = NULL;
4119 }
4120 mtx_unlock(&sync_mtx);
4121 BO_UNLOCK(bo);
4122 if (vp != NULL) {
4123 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4124 vgone(vp);
4125 vput(vp);
4126 }
4127 }
4128
4129 void
4130 vfs_deallocate_syncvnode(struct mount *mp)
4131 {
4132 struct vnode *vp;
4133
4134 mtx_lock(&sync_mtx);
4135 vp = mp->mnt_syncer;
4136 if (vp != NULL)
4137 mp->mnt_syncer = NULL;
4138 mtx_unlock(&sync_mtx);
4139 if (vp != NULL)
4140 vrele(vp);
4141 }
4142
4143 /*
4144 * Do a lazy sync of the filesystem.
4145 */
4146 static int
4147 sync_fsync(struct vop_fsync_args *ap)
4148 {
4149 struct vnode *syncvp = ap->a_vp;
4150 struct mount *mp = syncvp->v_mount;
4151 int error, save;
4152 struct bufobj *bo;
4153
4154 /*
4155 * We only need to do something if this is a lazy evaluation.
4156 */
4157 if (ap->a_waitfor != MNT_LAZY)
4158 return (0);
4159
4160 /*
4161 * Move ourselves to the back of the sync list.
4162 */
4163 bo = &syncvp->v_bufobj;
4164 BO_LOCK(bo);
4165 vn_syncer_add_to_worklist(bo, syncdelay);
4166 BO_UNLOCK(bo);
4167
4168 /*
4169 * Walk the list of vnodes pushing all that are dirty and
4170 * not already on the sync list.
4171 */
4172 if (vfs_busy(mp, MBF_NOWAIT) != 0)
4173 return (0);
4174 if (vn_start_write(NULL, &mp, V_NOWAIT) != 0) {
4175 vfs_unbusy(mp);
4176 return (0);
4177 }
4178 save = curthread_pflags_set(TDP_SYNCIO);
4179 vfs_msync(mp, MNT_NOWAIT);
4180 error = VFS_SYNC(mp, MNT_LAZY);
4181 curthread_pflags_restore(save);
4182 vn_finished_write(mp);
4183 vfs_unbusy(mp);
4184 return (error);
4185 }
4186
4187 /*
4188 * The syncer vnode is no referenced.
4189 */
4190 static int
4191 sync_inactive(struct vop_inactive_args *ap)
4192 {
4193
4194 vgone(ap->a_vp);
4195 return (0);
4196 }
4197
4198 /*
4199 * The syncer vnode is no longer needed and is being decommissioned.
4200 *
4201 * Modifications to the worklist must be protected by sync_mtx.
4202 */
4203 static int
4204 sync_reclaim(struct vop_reclaim_args *ap)
4205 {
4206 struct vnode *vp = ap->a_vp;
4207 struct bufobj *bo;
4208
4209 bo = &vp->v_bufobj;
4210 BO_LOCK(bo);
4211 mtx_lock(&sync_mtx);
4212 if (vp->v_mount->mnt_syncer == vp)
4213 vp->v_mount->mnt_syncer = NULL;
4214 if (bo->bo_flag & BO_ONWORKLST) {
4215 LIST_REMOVE(bo, bo_synclist);
4216 syncer_worklist_len--;
4217 sync_vnode_count--;
4218 bo->bo_flag &= ~BO_ONWORKLST;
4219 }
4220 mtx_unlock(&sync_mtx);
4221 BO_UNLOCK(bo);
4222
4223 return (0);
4224 }
4225
4226 /*
4227 * Check if vnode represents a disk device
4228 */
4229 int
4230 vn_isdisk(struct vnode *vp, int *errp)
4231 {
4232 int error;
4233
4234 if (vp->v_type != VCHR) {
4235 error = ENOTBLK;
4236 goto out;
4237 }
4238 error = 0;
4239 dev_lock();
4240 if (vp->v_rdev == NULL)
4241 error = ENXIO;
4242 else if (vp->v_rdev->si_devsw == NULL)
4243 error = ENXIO;
4244 else if (!(vp->v_rdev->si_devsw->d_flags & D_DISK))
4245 error = ENOTBLK;
4246 dev_unlock();
4247 out:
4248 if (errp != NULL)
4249 *errp = error;
4250 return (error == 0);
4251 }
4252
4253 /*
4254 * Common filesystem object access control check routine. Accepts a
4255 * vnode's type, "mode", uid and gid, requested access mode, credentials,
4256 * and optional call-by-reference privused argument allowing vaccess()
4257 * to indicate to the caller whether privilege was used to satisfy the
4258 * request (obsoleted). Returns 0 on success, or an errno on failure.
4259 */
4260 int
4261 vaccess(enum vtype type, mode_t file_mode, uid_t file_uid, gid_t file_gid,
4262 accmode_t accmode, struct ucred *cred, int *privused)
4263 {
4264 accmode_t dac_granted;
4265 accmode_t priv_granted;
4266
4267 KASSERT((accmode & ~(VEXEC | VWRITE | VREAD | VADMIN | VAPPEND)) == 0,
4268 ("invalid bit in accmode"));
4269 KASSERT((accmode & VAPPEND) == 0 || (accmode & VWRITE),
4270 ("VAPPEND without VWRITE"));
4271
4272 /*
4273 * Look for a normal, non-privileged way to access the file/directory
4274 * as requested. If it exists, go with that.
4275 */
4276
4277 if (privused != NULL)
4278 *privused = 0;
4279
4280 dac_granted = 0;
4281
4282 /* Check the owner. */
4283 if (cred->cr_uid == file_uid) {
4284 dac_granted |= VADMIN;
4285 if (file_mode & S_IXUSR)
4286 dac_granted |= VEXEC;
4287 if (file_mode & S_IRUSR)
4288 dac_granted |= VREAD;
4289 if (file_mode & S_IWUSR)
4290 dac_granted |= (VWRITE | VAPPEND);
4291
4292 if ((accmode & dac_granted) == accmode)
4293 return (0);
4294
4295 goto privcheck;
4296 }
4297
4298 /* Otherwise, check the groups (first match) */
4299 if (groupmember(file_gid, cred)) {
4300 if (file_mode & S_IXGRP)
4301 dac_granted |= VEXEC;
4302 if (file_mode & S_IRGRP)
4303 dac_granted |= VREAD;
4304 if (file_mode & S_IWGRP)
4305 dac_granted |= (VWRITE | VAPPEND);
4306
4307 if ((accmode & dac_granted) == accmode)
4308 return (0);
4309
4310 goto privcheck;
4311 }
4312
4313 /* Otherwise, check everyone else. */
4314 if (file_mode & S_IXOTH)
4315 dac_granted |= VEXEC;
4316 if (file_mode & S_IROTH)
4317 dac_granted |= VREAD;
4318 if (file_mode & S_IWOTH)
4319 dac_granted |= (VWRITE | VAPPEND);
4320 if ((accmode & dac_granted) == accmode)
4321 return (0);
4322
4323 privcheck:
4324 /*
4325 * Build a privilege mask to determine if the set of privileges
4326 * satisfies the requirements when combined with the granted mask
4327 * from above. For each privilege, if the privilege is required,
4328 * bitwise or the request type onto the priv_granted mask.
4329 */
4330 priv_granted = 0;
4331
4332 if (type == VDIR) {
4333 /*
4334 * For directories, use PRIV_VFS_LOOKUP to satisfy VEXEC
4335 * requests, instead of PRIV_VFS_EXEC.
4336 */
4337 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4338 !priv_check_cred(cred, PRIV_VFS_LOOKUP, 0))
4339 priv_granted |= VEXEC;
4340 } else {
4341 /*
4342 * Ensure that at least one execute bit is on. Otherwise,
4343 * a privileged user will always succeed, and we don't want
4344 * this to happen unless the file really is executable.
4345 */
4346 if ((accmode & VEXEC) && ((dac_granted & VEXEC) == 0) &&
4347 (file_mode & (S_IXUSR | S_IXGRP | S_IXOTH)) != 0 &&
4348 !priv_check_cred(cred, PRIV_VFS_EXEC, 0))
4349 priv_granted |= VEXEC;
4350 }
4351
4352 if ((accmode & VREAD) && ((dac_granted & VREAD) == 0) &&
4353 !priv_check_cred(cred, PRIV_VFS_READ, 0))
4354 priv_granted |= VREAD;
4355
4356 if ((accmode & VWRITE) && ((dac_granted & VWRITE) == 0) &&
4357 !priv_check_cred(cred, PRIV_VFS_WRITE, 0))
4358 priv_granted |= (VWRITE | VAPPEND);
4359
4360 if ((accmode & VADMIN) && ((dac_granted & VADMIN) == 0) &&
4361 !priv_check_cred(cred, PRIV_VFS_ADMIN, 0))
4362 priv_granted |= VADMIN;
4363
4364 if ((accmode & (priv_granted | dac_granted)) == accmode) {
4365 /* XXX audit: privilege used */
4366 if (privused != NULL)
4367 *privused = 1;
4368 return (0);
4369 }
4370
4371 return ((accmode & VADMIN) ? EPERM : EACCES);
4372 }
4373
4374 /*
4375 * Credential check based on process requesting service, and per-attribute
4376 * permissions.
4377 */
4378 int
4379 extattr_check_cred(struct vnode *vp, int attrnamespace, struct ucred *cred,
4380 struct thread *td, accmode_t accmode)
4381 {
4382
4383 /*
4384 * Kernel-invoked always succeeds.
4385 */
4386 if (cred == NOCRED)
4387 return (0);
4388
4389 /*
4390 * Do not allow privileged processes in jail to directly manipulate
4391 * system attributes.
4392 */
4393 switch (attrnamespace) {
4394 case EXTATTR_NAMESPACE_SYSTEM:
4395 /* Potentially should be: return (EPERM); */
4396 return (priv_check_cred(cred, PRIV_VFS_EXTATTR_SYSTEM, 0));
4397 case EXTATTR_NAMESPACE_USER:
4398 return (VOP_ACCESS(vp, accmode, cred, td));
4399 default:
4400 return (EPERM);
4401 }
4402 }
4403
4404 #ifdef DEBUG_VFS_LOCKS
4405 /*
4406 * This only exists to suppress warnings from unlocked specfs accesses. It is
4407 * no longer ok to have an unlocked VFS.
4408 */
4409 #define IGNORE_LOCK(vp) (panicstr != NULL || (vp) == NULL || \
4410 (vp)->v_type == VCHR || (vp)->v_type == VBAD)
4411
4412 int vfs_badlock_ddb = 1; /* Drop into debugger on violation. */
4413 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_ddb, CTLFLAG_RW, &vfs_badlock_ddb, 0,
4414 "Drop into debugger on lock violation");
4415
4416 int vfs_badlock_mutex = 1; /* Check for interlock across VOPs. */
4417 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_mutex, CTLFLAG_RW, &vfs_badlock_mutex,
4418 0, "Check for interlock across VOPs");
4419
4420 int vfs_badlock_print = 1; /* Print lock violations. */
4421 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_print, CTLFLAG_RW, &vfs_badlock_print,
4422 0, "Print lock violations");
4423
4424 int vfs_badlock_vnode = 1; /* Print vnode details on lock violations. */
4425 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_vnode, CTLFLAG_RW, &vfs_badlock_vnode,
4426 0, "Print vnode details on lock violations");
4427
4428 #ifdef KDB
4429 int vfs_badlock_backtrace = 1; /* Print backtrace at lock violations. */
4430 SYSCTL_INT(_debug, OID_AUTO, vfs_badlock_backtrace, CTLFLAG_RW,
4431 &vfs_badlock_backtrace, 0, "Print backtrace at lock violations");
4432 #endif
4433
4434 static void
4435 vfs_badlock(const char *msg, const char *str, struct vnode *vp)
4436 {
4437
4438 #ifdef KDB
4439 if (vfs_badlock_backtrace)
4440 kdb_backtrace();
4441 #endif
4442 if (vfs_badlock_vnode)
4443 vn_printf(vp, "vnode ");
4444 if (vfs_badlock_print)
4445 printf("%s: %p %s\n", str, (void *)vp, msg);
4446 if (vfs_badlock_ddb)
4447 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4448 }
4449
4450 void
4451 assert_vi_locked(struct vnode *vp, const char *str)
4452 {
4453
4454 if (vfs_badlock_mutex && !mtx_owned(VI_MTX(vp)))
4455 vfs_badlock("interlock is not locked but should be", str, vp);
4456 }
4457
4458 void
4459 assert_vi_unlocked(struct vnode *vp, const char *str)
4460 {
4461
4462 if (vfs_badlock_mutex && mtx_owned(VI_MTX(vp)))
4463 vfs_badlock("interlock is locked but should not be", str, vp);
4464 }
4465
4466 void
4467 assert_vop_locked(struct vnode *vp, const char *str)
4468 {
4469 int locked;
4470
4471 if (!IGNORE_LOCK(vp)) {
4472 locked = VOP_ISLOCKED(vp);
4473 if (locked == 0 || locked == LK_EXCLOTHER)
4474 vfs_badlock("is not locked but should be", str, vp);
4475 }
4476 }
4477
4478 void
4479 assert_vop_unlocked(struct vnode *vp, const char *str)
4480 {
4481
4482 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) == LK_EXCLUSIVE)
4483 vfs_badlock("is locked but should not be", str, vp);
4484 }
4485
4486 void
4487 assert_vop_elocked(struct vnode *vp, const char *str)
4488 {
4489
4490 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLUSIVE)
4491 vfs_badlock("is not exclusive locked but should be", str, vp);
4492 }
4493
4494 #if 0
4495 void
4496 assert_vop_elocked_other(struct vnode *vp, const char *str)
4497 {
4498
4499 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_EXCLOTHER)
4500 vfs_badlock("is not exclusive locked by another thread",
4501 str, vp);
4502 }
4503
4504 void
4505 assert_vop_slocked(struct vnode *vp, const char *str)
4506 {
4507
4508 if (!IGNORE_LOCK(vp) && VOP_ISLOCKED(vp) != LK_SHARED)
4509 vfs_badlock("is not locked shared but should be", str, vp);
4510 }
4511 #endif /* 0 */
4512 #endif /* DEBUG_VFS_LOCKS */
4513
4514 void
4515 vop_rename_fail(struct vop_rename_args *ap)
4516 {
4517
4518 if (ap->a_tvp != NULL)
4519 vput(ap->a_tvp);
4520 if (ap->a_tdvp == ap->a_tvp)
4521 vrele(ap->a_tdvp);
4522 else
4523 vput(ap->a_tdvp);
4524 vrele(ap->a_fdvp);
4525 vrele(ap->a_fvp);
4526 }
4527
4528 void
4529 vop_rename_pre(void *ap)
4530 {
4531 struct vop_rename_args *a = ap;
4532
4533 #ifdef DEBUG_VFS_LOCKS
4534 if (a->a_tvp)
4535 ASSERT_VI_UNLOCKED(a->a_tvp, "VOP_RENAME");
4536 ASSERT_VI_UNLOCKED(a->a_tdvp, "VOP_RENAME");
4537 ASSERT_VI_UNLOCKED(a->a_fvp, "VOP_RENAME");
4538 ASSERT_VI_UNLOCKED(a->a_fdvp, "VOP_RENAME");
4539
4540 /* Check the source (from). */
4541 if (a->a_tdvp->v_vnlock != a->a_fdvp->v_vnlock &&
4542 (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fdvp->v_vnlock))
4543 ASSERT_VOP_UNLOCKED(a->a_fdvp, "vop_rename: fdvp locked");
4544 if (a->a_tvp == NULL || a->a_tvp->v_vnlock != a->a_fvp->v_vnlock)
4545 ASSERT_VOP_UNLOCKED(a->a_fvp, "vop_rename: fvp locked");
4546
4547 /* Check the target. */
4548 if (a->a_tvp)
4549 ASSERT_VOP_LOCKED(a->a_tvp, "vop_rename: tvp not locked");
4550 ASSERT_VOP_LOCKED(a->a_tdvp, "vop_rename: tdvp not locked");
4551 #endif
4552 if (a->a_tdvp != a->a_fdvp)
4553 vhold(a->a_fdvp);
4554 if (a->a_tvp != a->a_fvp)
4555 vhold(a->a_fvp);
4556 vhold(a->a_tdvp);
4557 if (a->a_tvp)
4558 vhold(a->a_tvp);
4559 }
4560
4561 #ifdef DEBUG_VFS_LOCKS
4562 void
4563 vop_strategy_pre(void *ap)
4564 {
4565 struct vop_strategy_args *a;
4566 struct buf *bp;
4567
4568 a = ap;
4569 bp = a->a_bp;
4570
4571 /*
4572 * Cluster ops lock their component buffers but not the IO container.
4573 */
4574 if ((bp->b_flags & B_CLUSTER) != 0)
4575 return;
4576
4577 if (panicstr == NULL && !BUF_ISLOCKED(bp)) {
4578 if (vfs_badlock_print)
4579 printf(
4580 "VOP_STRATEGY: bp is not locked but should be\n");
4581 if (vfs_badlock_ddb)
4582 kdb_enter(KDB_WHY_VFSLOCK, "lock violation");
4583 }
4584 }
4585
4586 void
4587 vop_lock_pre(void *ap)
4588 {
4589 struct vop_lock1_args *a = ap;
4590
4591 if ((a->a_flags & LK_INTERLOCK) == 0)
4592 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4593 else
4594 ASSERT_VI_LOCKED(a->a_vp, "VOP_LOCK");
4595 }
4596
4597 void
4598 vop_lock_post(void *ap, int rc)
4599 {
4600 struct vop_lock1_args *a = ap;
4601
4602 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_LOCK");
4603 if (rc == 0 && (a->a_flags & LK_EXCLOTHER) == 0)
4604 ASSERT_VOP_LOCKED(a->a_vp, "VOP_LOCK");
4605 }
4606
4607 void
4608 vop_unlock_pre(void *ap)
4609 {
4610 struct vop_unlock_args *a = ap;
4611
4612 if (a->a_flags & LK_INTERLOCK)
4613 ASSERT_VI_LOCKED(a->a_vp, "VOP_UNLOCK");
4614 ASSERT_VOP_LOCKED(a->a_vp, "VOP_UNLOCK");
4615 }
4616
4617 void
4618 vop_unlock_post(void *ap, int rc)
4619 {
4620 struct vop_unlock_args *a = ap;
4621
4622 if (a->a_flags & LK_INTERLOCK)
4623 ASSERT_VI_UNLOCKED(a->a_vp, "VOP_UNLOCK");
4624 }
4625 #endif
4626
4627 void
4628 vop_create_post(void *ap, int rc)
4629 {
4630 struct vop_create_args *a = ap;
4631
4632 if (!rc)
4633 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4634 }
4635
4636 void
4637 vop_deleteextattr_post(void *ap, int rc)
4638 {
4639 struct vop_deleteextattr_args *a = ap;
4640
4641 if (!rc)
4642 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4643 }
4644
4645 void
4646 vop_link_post(void *ap, int rc)
4647 {
4648 struct vop_link_args *a = ap;
4649
4650 if (!rc) {
4651 VFS_KNOTE_LOCKED(a->a_vp, NOTE_LINK);
4652 VFS_KNOTE_LOCKED(a->a_tdvp, NOTE_WRITE);
4653 }
4654 }
4655
4656 void
4657 vop_mkdir_post(void *ap, int rc)
4658 {
4659 struct vop_mkdir_args *a = ap;
4660
4661 if (!rc)
4662 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4663 }
4664
4665 void
4666 vop_mknod_post(void *ap, int rc)
4667 {
4668 struct vop_mknod_args *a = ap;
4669
4670 if (!rc)
4671 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4672 }
4673
4674 void
4675 vop_reclaim_post(void *ap, int rc)
4676 {
4677 struct vop_reclaim_args *a = ap;
4678
4679 if (!rc)
4680 VFS_KNOTE_LOCKED(a->a_vp, NOTE_REVOKE);
4681 }
4682
4683 void
4684 vop_remove_post(void *ap, int rc)
4685 {
4686 struct vop_remove_args *a = ap;
4687
4688 if (!rc) {
4689 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4690 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4691 }
4692 }
4693
4694 void
4695 vop_rename_post(void *ap, int rc)
4696 {
4697 struct vop_rename_args *a = ap;
4698 long hint;
4699
4700 if (!rc) {
4701 hint = NOTE_WRITE;
4702 if (a->a_fdvp == a->a_tdvp) {
4703 if (a->a_tvp != NULL && a->a_tvp->v_type == VDIR)
4704 hint |= NOTE_LINK;
4705 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4706 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4707 } else {
4708 hint |= NOTE_EXTEND;
4709 if (a->a_fvp->v_type == VDIR)
4710 hint |= NOTE_LINK;
4711 VFS_KNOTE_UNLOCKED(a->a_fdvp, hint);
4712
4713 if (a->a_fvp->v_type == VDIR && a->a_tvp != NULL &&
4714 a->a_tvp->v_type == VDIR)
4715 hint &= ~NOTE_LINK;
4716 VFS_KNOTE_UNLOCKED(a->a_tdvp, hint);
4717 }
4718
4719 VFS_KNOTE_UNLOCKED(a->a_fvp, NOTE_RENAME);
4720 if (a->a_tvp)
4721 VFS_KNOTE_UNLOCKED(a->a_tvp, NOTE_DELETE);
4722 }
4723 if (a->a_tdvp != a->a_fdvp)
4724 vdrop(a->a_fdvp);
4725 if (a->a_tvp != a->a_fvp)
4726 vdrop(a->a_fvp);
4727 vdrop(a->a_tdvp);
4728 if (a->a_tvp)
4729 vdrop(a->a_tvp);
4730 }
4731
4732 void
4733 vop_rmdir_post(void *ap, int rc)
4734 {
4735 struct vop_rmdir_args *a = ap;
4736
4737 if (!rc) {
4738 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE | NOTE_LINK);
4739 VFS_KNOTE_LOCKED(a->a_vp, NOTE_DELETE);
4740 }
4741 }
4742
4743 void
4744 vop_setattr_post(void *ap, int rc)
4745 {
4746 struct vop_setattr_args *a = ap;
4747
4748 if (!rc)
4749 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4750 }
4751
4752 void
4753 vop_setextattr_post(void *ap, int rc)
4754 {
4755 struct vop_setextattr_args *a = ap;
4756
4757 if (!rc)
4758 VFS_KNOTE_LOCKED(a->a_vp, NOTE_ATTRIB);
4759 }
4760
4761 void
4762 vop_symlink_post(void *ap, int rc)
4763 {
4764 struct vop_symlink_args *a = ap;
4765
4766 if (!rc)
4767 VFS_KNOTE_LOCKED(a->a_dvp, NOTE_WRITE);
4768 }
4769
4770 void
4771 vop_open_post(void *ap, int rc)
4772 {
4773 struct vop_open_args *a = ap;
4774
4775 if (!rc)
4776 VFS_KNOTE_LOCKED(a->a_vp, NOTE_OPEN);
4777 }
4778
4779 void
4780 vop_close_post(void *ap, int rc)
4781 {
4782 struct vop_close_args *a = ap;
4783
4784 if (!rc && (a->a_cred != NOCRED || /* filter out revokes */
4785 (a->a_vp->v_iflag & VI_DOOMED) == 0)) {
4786 VFS_KNOTE_LOCKED(a->a_vp, (a->a_fflag & FWRITE) != 0 ?
4787 NOTE_CLOSE_WRITE : NOTE_CLOSE);
4788 }
4789 }
4790
4791 void
4792 vop_read_post(void *ap, int rc)
4793 {
4794 struct vop_read_args *a = ap;
4795
4796 if (!rc)
4797 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
4798 }
4799
4800 void
4801 vop_readdir_post(void *ap, int rc)
4802 {
4803 struct vop_readdir_args *a = ap;
4804
4805 if (!rc)
4806 VFS_KNOTE_LOCKED(a->a_vp, NOTE_READ);
4807 }
4808
4809 static struct knlist fs_knlist;
4810
4811 static void
4812 vfs_event_init(void *arg)
4813 {
4814 knlist_init_mtx(&fs_knlist, NULL);
4815 }
4816 /* XXX - correct order? */
4817 SYSINIT(vfs_knlist, SI_SUB_VFS, SI_ORDER_ANY, vfs_event_init, NULL);
4818
4819 void
4820 vfs_event_signal(fsid_t *fsid, uint32_t event, intptr_t data __unused)
4821 {
4822
4823 KNOTE_UNLOCKED(&fs_knlist, event);
4824 }
4825
4826 static int filt_fsattach(struct knote *kn);
4827 static void filt_fsdetach(struct knote *kn);
4828 static int filt_fsevent(struct knote *kn, long hint);
4829
4830 struct filterops fs_filtops = {
4831 .f_isfd = 0,
4832 .f_attach = filt_fsattach,
4833 .f_detach = filt_fsdetach,
4834 .f_event = filt_fsevent
4835 };
4836
4837 static int
4838 filt_fsattach(struct knote *kn)
4839 {
4840
4841 kn->kn_flags |= EV_CLEAR;
4842 knlist_add(&fs_knlist, kn, 0);
4843 return (0);
4844 }
4845
4846 static void
4847 filt_fsdetach(struct knote *kn)
4848 {
4849
4850 knlist_remove(&fs_knlist, kn, 0);
4851 }
4852
4853 static int
4854 filt_fsevent(struct knote *kn, long hint)
4855 {
4856
4857 kn->kn_fflags |= hint;
4858 return (kn->kn_fflags != 0);
4859 }
4860
4861 static int
4862 sysctl_vfs_ctl(SYSCTL_HANDLER_ARGS)
4863 {
4864 struct vfsidctl vc;
4865 int error;
4866 struct mount *mp;
4867
4868 error = SYSCTL_IN(req, &vc, sizeof(vc));
4869 if (error)
4870 return (error);
4871 if (vc.vc_vers != VFS_CTL_VERS1)
4872 return (EINVAL);
4873 mp = vfs_getvfs(&vc.vc_fsid);
4874 if (mp == NULL)
4875 return (ENOENT);
4876 /* ensure that a specific sysctl goes to the right filesystem. */
4877 if (strcmp(vc.vc_fstypename, "*") != 0 &&
4878 strcmp(vc.vc_fstypename, mp->mnt_vfc->vfc_name) != 0) {
4879 vfs_rel(mp);
4880 return (EINVAL);
4881 }
4882 VCTLTOREQ(&vc, req);
4883 error = VFS_SYSCTL(mp, vc.vc_op, req);
4884 vfs_rel(mp);
4885 return (error);
4886 }
4887
4888 SYSCTL_PROC(_vfs, OID_AUTO, ctl, CTLTYPE_OPAQUE | CTLFLAG_WR,
4889 NULL, 0, sysctl_vfs_ctl, "",
4890 "Sysctl by fsid");
4891
4892 /*
4893 * Function to initialize a va_filerev field sensibly.
4894 * XXX: Wouldn't a random number make a lot more sense ??
4895 */
4896 u_quad_t
4897 init_va_filerev(void)
4898 {
4899 struct bintime bt;
4900
4901 getbinuptime(&bt);
4902 return (((u_quad_t)bt.sec << 32LL) | (bt.frac >> 32LL));
4903 }
4904
4905 static int filt_vfsread(struct knote *kn, long hint);
4906 static int filt_vfswrite(struct knote *kn, long hint);
4907 static int filt_vfsvnode(struct knote *kn, long hint);
4908 static void filt_vfsdetach(struct knote *kn);
4909 static struct filterops vfsread_filtops = {
4910 .f_isfd = 1,
4911 .f_detach = filt_vfsdetach,
4912 .f_event = filt_vfsread
4913 };
4914 static struct filterops vfswrite_filtops = {
4915 .f_isfd = 1,
4916 .f_detach = filt_vfsdetach,
4917 .f_event = filt_vfswrite
4918 };
4919 static struct filterops vfsvnode_filtops = {
4920 .f_isfd = 1,
4921 .f_detach = filt_vfsdetach,
4922 .f_event = filt_vfsvnode
4923 };
4924
4925 static void
4926 vfs_knllock(void *arg)
4927 {
4928 struct vnode *vp = arg;
4929
4930 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
4931 }
4932
4933 static void
4934 vfs_knlunlock(void *arg)
4935 {
4936 struct vnode *vp = arg;
4937
4938 VOP_UNLOCK(vp, 0);
4939 }
4940
4941 static void
4942 vfs_knl_assert_locked(void *arg)
4943 {
4944 #ifdef DEBUG_VFS_LOCKS
4945 struct vnode *vp = arg;
4946
4947 ASSERT_VOP_LOCKED(vp, "vfs_knl_assert_locked");
4948 #endif
4949 }
4950
4951 static void
4952 vfs_knl_assert_unlocked(void *arg)
4953 {
4954 #ifdef DEBUG_VFS_LOCKS
4955 struct vnode *vp = arg;
4956
4957 ASSERT_VOP_UNLOCKED(vp, "vfs_knl_assert_unlocked");
4958 #endif
4959 }
4960
4961 int
4962 vfs_kqfilter(struct vop_kqfilter_args *ap)
4963 {
4964 struct vnode *vp = ap->a_vp;
4965 struct knote *kn = ap->a_kn;
4966 struct knlist *knl;
4967
4968 switch (kn->kn_filter) {
4969 case EVFILT_READ:
4970 kn->kn_fop = &vfsread_filtops;
4971 break;
4972 case EVFILT_WRITE:
4973 kn->kn_fop = &vfswrite_filtops;
4974 break;
4975 case EVFILT_VNODE:
4976 kn->kn_fop = &vfsvnode_filtops;
4977 break;
4978 default:
4979 return (EINVAL);
4980 }
4981
4982 kn->kn_hook = (caddr_t)vp;
4983
4984 v_addpollinfo(vp);
4985 if (vp->v_pollinfo == NULL)
4986 return (ENOMEM);
4987 knl = &vp->v_pollinfo->vpi_selinfo.si_note;
4988 vhold(vp);
4989 knlist_add(knl, kn, 0);
4990
4991 return (0);
4992 }
4993
4994 /*
4995 * Detach knote from vnode
4996 */
4997 static void
4998 filt_vfsdetach(struct knote *kn)
4999 {
5000 struct vnode *vp = (struct vnode *)kn->kn_hook;
5001
5002 KASSERT(vp->v_pollinfo != NULL, ("Missing v_pollinfo"));
5003 knlist_remove(&vp->v_pollinfo->vpi_selinfo.si_note, kn, 0);
5004 vdrop(vp);
5005 }
5006
5007 /*ARGSUSED*/
5008 static int
5009 filt_vfsread(struct knote *kn, long hint)
5010 {
5011 struct vnode *vp = (struct vnode *)kn->kn_hook;
5012 struct vattr va;
5013 int res;
5014
5015 /*
5016 * filesystem is gone, so set the EOF flag and schedule
5017 * the knote for deletion.
5018 */
5019 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5020 VI_LOCK(vp);
5021 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5022 VI_UNLOCK(vp);
5023 return (1);
5024 }
5025
5026 if (VOP_GETATTR(vp, &va, curthread->td_ucred))
5027 return (0);
5028
5029 VI_LOCK(vp);
5030 kn->kn_data = va.va_size - kn->kn_fp->f_offset;
5031 res = (kn->kn_sfflags & NOTE_FILE_POLL) != 0 || kn->kn_data != 0;
5032 VI_UNLOCK(vp);
5033 return (res);
5034 }
5035
5036 /*ARGSUSED*/
5037 static int
5038 filt_vfswrite(struct knote *kn, long hint)
5039 {
5040 struct vnode *vp = (struct vnode *)kn->kn_hook;
5041
5042 VI_LOCK(vp);
5043
5044 /*
5045 * filesystem is gone, so set the EOF flag and schedule
5046 * the knote for deletion.
5047 */
5048 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD))
5049 kn->kn_flags |= (EV_EOF | EV_ONESHOT);
5050
5051 kn->kn_data = 0;
5052 VI_UNLOCK(vp);
5053 return (1);
5054 }
5055
5056 static int
5057 filt_vfsvnode(struct knote *kn, long hint)
5058 {
5059 struct vnode *vp = (struct vnode *)kn->kn_hook;
5060 int res;
5061
5062 VI_LOCK(vp);
5063 if (kn->kn_sfflags & hint)
5064 kn->kn_fflags |= hint;
5065 if (hint == NOTE_REVOKE || (hint == 0 && vp->v_type == VBAD)) {
5066 kn->kn_flags |= EV_EOF;
5067 VI_UNLOCK(vp);
5068 return (1);
5069 }
5070 res = (kn->kn_fflags != 0);
5071 VI_UNLOCK(vp);
5072 return (res);
5073 }
5074
5075 int
5076 vfs_read_dirent(struct vop_readdir_args *ap, struct dirent *dp, off_t off)
5077 {
5078 int error;
5079
5080 if (dp->d_reclen > ap->a_uio->uio_resid)
5081 return (ENAMETOOLONG);
5082 error = uiomove(dp, dp->d_reclen, ap->a_uio);
5083 if (error) {
5084 if (ap->a_ncookies != NULL) {
5085 if (ap->a_cookies != NULL)
5086 free(ap->a_cookies, M_TEMP);
5087 ap->a_cookies = NULL;
5088 *ap->a_ncookies = 0;
5089 }
5090 return (error);
5091 }
5092 if (ap->a_ncookies == NULL)
5093 return (0);
5094
5095 KASSERT(ap->a_cookies,
5096 ("NULL ap->a_cookies value with non-NULL ap->a_ncookies!"));
5097
5098 *ap->a_cookies = realloc(*ap->a_cookies,
5099 (*ap->a_ncookies + 1) * sizeof(u_long), M_TEMP, M_WAITOK | M_ZERO);
5100 (*ap->a_cookies)[*ap->a_ncookies] = off;
5101 *ap->a_ncookies += 1;
5102 return (0);
5103 }
5104
5105 /*
5106 * Mark for update the access time of the file if the filesystem
5107 * supports VOP_MARKATIME. This functionality is used by execve and
5108 * mmap, so we want to avoid the I/O implied by directly setting
5109 * va_atime for the sake of efficiency.
5110 */
5111 void
5112 vfs_mark_atime(struct vnode *vp, struct ucred *cred)
5113 {
5114 struct mount *mp;
5115
5116 mp = vp->v_mount;
5117 ASSERT_VOP_LOCKED(vp, "vfs_mark_atime");
5118 if (mp != NULL && (mp->mnt_flag & (MNT_NOATIME | MNT_RDONLY)) == 0)
5119 (void)VOP_MARKATIME(vp);
5120 }
5121
5122 /*
5123 * The purpose of this routine is to remove granularity from accmode_t,
5124 * reducing it into standard unix access bits - VEXEC, VREAD, VWRITE,
5125 * VADMIN and VAPPEND.
5126 *
5127 * If it returns 0, the caller is supposed to continue with the usual
5128 * access checks using 'accmode' as modified by this routine. If it
5129 * returns nonzero value, the caller is supposed to return that value
5130 * as errno.
5131 *
5132 * Note that after this routine runs, accmode may be zero.
5133 */
5134 int
5135 vfs_unixify_accmode(accmode_t *accmode)
5136 {
5137 /*
5138 * There is no way to specify explicit "deny" rule using
5139 * file mode or POSIX.1e ACLs.
5140 */
5141 if (*accmode & VEXPLICIT_DENY) {
5142 *accmode = 0;
5143 return (0);
5144 }
5145
5146 /*
5147 * None of these can be translated into usual access bits.
5148 * Also, the common case for NFSv4 ACLs is to not contain
5149 * either of these bits. Caller should check for VWRITE
5150 * on the containing directory instead.
5151 */
5152 if (*accmode & (VDELETE_CHILD | VDELETE))
5153 return (EPERM);
5154
5155 if (*accmode & VADMIN_PERMS) {
5156 *accmode &= ~VADMIN_PERMS;
5157 *accmode |= VADMIN;
5158 }
5159
5160 /*
5161 * There is no way to deny VREAD_ATTRIBUTES, VREAD_ACL
5162 * or VSYNCHRONIZE using file mode or POSIX.1e ACL.
5163 */
5164 *accmode &= ~(VSTAT_PERMS | VSYNCHRONIZE);
5165
5166 return (0);
5167 }
5168
5169 /*
5170 * These are helper functions for filesystems to traverse all
5171 * their vnodes. See MNT_VNODE_FOREACH_ALL() in sys/mount.h.
5172 *
5173 * This interface replaces MNT_VNODE_FOREACH.
5174 */
5175
5176 MALLOC_DEFINE(M_VNODE_MARKER, "vnodemarker", "vnode marker");
5177
5178 struct vnode *
5179 __mnt_vnode_next_all(struct vnode **mvp, struct mount *mp)
5180 {
5181 struct vnode *vp;
5182
5183 if (should_yield())
5184 kern_yield(PRI_USER);
5185 MNT_ILOCK(mp);
5186 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5187 for (vp = TAILQ_NEXT(*mvp, v_nmntvnodes); vp != NULL;
5188 vp = TAILQ_NEXT(vp, v_nmntvnodes)) {
5189 /* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5190 if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5191 continue;
5192 VI_LOCK(vp);
5193 if ((vp->v_iflag & VI_DOOMED) != 0) {
5194 VI_UNLOCK(vp);
5195 continue;
5196 }
5197 break;
5198 }
5199 if (vp == NULL) {
5200 __mnt_vnode_markerfree_all(mvp, mp);
5201 /* MNT_IUNLOCK(mp); -- done in above function */
5202 mtx_assert(MNT_MTX(mp), MA_NOTOWNED);
5203 return (NULL);
5204 }
5205 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5206 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5207 MNT_IUNLOCK(mp);
5208 return (vp);
5209 }
5210
5211 struct vnode *
5212 __mnt_vnode_first_all(struct vnode **mvp, struct mount *mp)
5213 {
5214 struct vnode *vp;
5215
5216 *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5217 MNT_ILOCK(mp);
5218 MNT_REF(mp);
5219 (*mvp)->v_mount = mp;
5220 (*mvp)->v_type = VMARKER;
5221
5222 TAILQ_FOREACH(vp, &mp->mnt_nvnodelist, v_nmntvnodes) {
5223 /* Allow a racy peek at VI_DOOMED to save a lock acquisition. */
5224 if (vp->v_type == VMARKER || (vp->v_iflag & VI_DOOMED) != 0)
5225 continue;
5226 VI_LOCK(vp);
5227 if ((vp->v_iflag & VI_DOOMED) != 0) {
5228 VI_UNLOCK(vp);
5229 continue;
5230 }
5231 break;
5232 }
5233 if (vp == NULL) {
5234 MNT_REL(mp);
5235 MNT_IUNLOCK(mp);
5236 free(*mvp, M_VNODE_MARKER);
5237 *mvp = NULL;
5238 return (NULL);
5239 }
5240 TAILQ_INSERT_AFTER(&mp->mnt_nvnodelist, vp, *mvp, v_nmntvnodes);
5241 MNT_IUNLOCK(mp);
5242 return (vp);
5243 }
5244
5245 void
5246 __mnt_vnode_markerfree_all(struct vnode **mvp, struct mount *mp)
5247 {
5248
5249 if (*mvp == NULL) {
5250 MNT_IUNLOCK(mp);
5251 return;
5252 }
5253
5254 mtx_assert(MNT_MTX(mp), MA_OWNED);
5255
5256 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5257 TAILQ_REMOVE(&mp->mnt_nvnodelist, *mvp, v_nmntvnodes);
5258 MNT_REL(mp);
5259 MNT_IUNLOCK(mp);
5260 free(*mvp, M_VNODE_MARKER);
5261 *mvp = NULL;
5262 }
5263
5264 /*
5265 * These are helper functions for filesystems to traverse their
5266 * active vnodes. See MNT_VNODE_FOREACH_ACTIVE() in sys/mount.h
5267 */
5268 static void
5269 mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5270 {
5271
5272 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5273
5274 MNT_ILOCK(mp);
5275 MNT_REL(mp);
5276 MNT_IUNLOCK(mp);
5277 free(*mvp, M_VNODE_MARKER);
5278 *mvp = NULL;
5279 }
5280
5281 static struct vnode *
5282 mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5283 {
5284 struct vnode *vp, *nvp;
5285
5286 mtx_assert(&vnode_free_list_mtx, MA_OWNED);
5287 KASSERT((*mvp)->v_mount == mp, ("marker vnode mount list mismatch"));
5288 restart:
5289 vp = TAILQ_NEXT(*mvp, v_actfreelist);
5290 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5291 while (vp != NULL) {
5292 if (vp->v_type == VMARKER) {
5293 vp = TAILQ_NEXT(vp, v_actfreelist);
5294 continue;
5295 }
5296 if (!VI_TRYLOCK(vp)) {
5297 if (mp_ncpus == 1 || should_yield()) {
5298 TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5299 mtx_unlock(&vnode_free_list_mtx);
5300 pause("vnacti", 1);
5301 mtx_lock(&vnode_free_list_mtx);
5302 goto restart;
5303 }
5304 continue;
5305 }
5306 KASSERT(vp->v_type != VMARKER, ("locked marker %p", vp));
5307 KASSERT(vp->v_mount == mp || vp->v_mount == NULL,
5308 ("alien vnode on the active list %p %p", vp, mp));
5309 if (vp->v_mount == mp && (vp->v_iflag & VI_DOOMED) == 0)
5310 break;
5311 nvp = TAILQ_NEXT(vp, v_actfreelist);
5312 VI_UNLOCK(vp);
5313 vp = nvp;
5314 }
5315
5316 /* Check if we are done */
5317 if (vp == NULL) {
5318 mtx_unlock(&vnode_free_list_mtx);
5319 mnt_vnode_markerfree_active(mvp, mp);
5320 return (NULL);
5321 }
5322 TAILQ_INSERT_AFTER(&mp->mnt_activevnodelist, vp, *mvp, v_actfreelist);
5323 mtx_unlock(&vnode_free_list_mtx);
5324 ASSERT_VI_LOCKED(vp, "active iter");
5325 KASSERT((vp->v_iflag & VI_ACTIVE) != 0, ("Non-active vp %p", vp));
5326 return (vp);
5327 }
5328
5329 struct vnode *
5330 __mnt_vnode_next_active(struct vnode **mvp, struct mount *mp)
5331 {
5332
5333 if (should_yield())
5334 kern_yield(PRI_USER);
5335 mtx_lock(&vnode_free_list_mtx);
5336 return (mnt_vnode_next_active(mvp, mp));
5337 }
5338
5339 struct vnode *
5340 __mnt_vnode_first_active(struct vnode **mvp, struct mount *mp)
5341 {
5342 struct vnode *vp;
5343
5344 *mvp = malloc(sizeof(struct vnode), M_VNODE_MARKER, M_WAITOK | M_ZERO);
5345 MNT_ILOCK(mp);
5346 MNT_REF(mp);
5347 MNT_IUNLOCK(mp);
5348 (*mvp)->v_type = VMARKER;
5349 (*mvp)->v_mount = mp;
5350
5351 mtx_lock(&vnode_free_list_mtx);
5352 vp = TAILQ_FIRST(&mp->mnt_activevnodelist);
5353 if (vp == NULL) {
5354 mtx_unlock(&vnode_free_list_mtx);
5355 mnt_vnode_markerfree_active(mvp, mp);
5356 return (NULL);
5357 }
5358 TAILQ_INSERT_BEFORE(vp, *mvp, v_actfreelist);
5359 return (mnt_vnode_next_active(mvp, mp));
5360 }
5361
5362 void
5363 __mnt_vnode_markerfree_active(struct vnode **mvp, struct mount *mp)
5364 {
5365
5366 if (*mvp == NULL)
5367 return;
5368
5369 mtx_lock(&vnode_free_list_mtx);
5370 TAILQ_REMOVE(&mp->mnt_activevnodelist, *mvp, v_actfreelist);
5371 mtx_unlock(&vnode_free_list_mtx);
5372 mnt_vnode_markerfree_active(mvp, mp);
5373 }
Cache object: 21532201899acd79b861d2601856b095
|