FreeBSD/Linux Kernel Cross Reference
sys/kern/vfs_vnode.c
1 /* $NetBSD: vfs_vnode.c,v 1.147 2022/10/26 23:40:08 riastradh Exp $ */
2
3 /*-
4 * Copyright (c) 1997-2011, 2019, 2020 The NetBSD Foundation, Inc.
5 * All rights reserved.
6 *
7 * This code is derived from software contributed to The NetBSD Foundation
8 * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
9 * NASA Ames Research Center, by Charles M. Hannum, and by Andrew Doran.
10 *
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 *
20 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
21 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
22 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
23 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
24 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
25 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
26 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
27 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
29 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
30 * POSSIBILITY OF SUCH DAMAGE.
31 */
32
33 /*
34 * Copyright (c) 1989, 1993
35 * The Regents of the University of California. All rights reserved.
36 * (c) UNIX System Laboratories, Inc.
37 * All or some portions of this file are derived from material licensed
38 * to the University of California by American Telephone and Telegraph
39 * Co. or Unix System Laboratories, Inc. and are reproduced herein with
40 * the permission of UNIX System Laboratories, Inc.
41 *
42 * Redistribution and use in source and binary forms, with or without
43 * modification, are permitted provided that the following conditions
44 * are met:
45 * 1. Redistributions of source code must retain the above copyright
46 * notice, this list of conditions and the following disclaimer.
47 * 2. Redistributions in binary form must reproduce the above copyright
48 * notice, this list of conditions and the following disclaimer in the
49 * documentation and/or other materials provided with the distribution.
50 * 3. Neither the name of the University nor the names of its contributors
51 * may be used to endorse or promote products derived from this software
52 * without specific prior written permission.
53 *
54 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64 * SUCH DAMAGE.
65 *
66 * @(#)vfs_subr.c 8.13 (Berkeley) 4/18/94
67 */
68
69 /*
70 * The vnode cache subsystem.
71 *
72 * Life-cycle
73 *
74 * Normally, there are two points where new vnodes are created:
75 * VOP_CREATE(9) and VOP_LOOKUP(9). The life-cycle of a vnode
76 * starts in one of the following ways:
77 *
78 * - Allocation, via vcache_get(9) or vcache_new(9).
79 * - Reclamation of inactive vnode, via vcache_vget(9).
80 *
81 * Recycle from a free list, via getnewvnode(9) -> getcleanvnode(9)
82 * was another, traditional way. Currently, only the draining thread
83 * recycles the vnodes. This behaviour might be revisited.
84 *
85 * The life-cycle ends when the last reference is dropped, usually
86 * in VOP_REMOVE(9). In such case, VOP_INACTIVE(9) is called to inform
87 * the file system that vnode is inactive. Via this call, file system
88 * indicates whether vnode can be recycled (usually, it checks its own
89 * references, e.g. count of links, whether the file was removed).
90 *
91 * Depending on indication, vnode can be put into a free list (cache),
92 * or cleaned via vcache_reclaim, which calls VOP_RECLAIM(9) to
93 * disassociate underlying file system from the vnode, and finally
94 * destroyed.
95 *
96 * Vnode state
97 *
98 * Vnode is always in one of six states:
99 * - MARKER This is a marker vnode to help list traversal. It
100 * will never change its state.
101 * - LOADING Vnode is associating underlying file system and not
102 * yet ready to use.
103 * - LOADED Vnode has associated underlying file system and is
104 * ready to use.
105 * - BLOCKED Vnode is active but cannot get new references.
106 * - RECLAIMING Vnode is disassociating from the underlying file
107 * system.
108 * - RECLAIMED Vnode has disassociated from underlying file system
109 * and is dead.
110 *
111 * Valid state changes are:
112 * LOADING -> LOADED
113 * Vnode has been initialised in vcache_get() or
114 * vcache_new() and is ready to use.
115 * BLOCKED -> RECLAIMING
116 * Vnode starts disassociation from underlying file
117 * system in vcache_reclaim().
118 * RECLAIMING -> RECLAIMED
119 * Vnode finished disassociation from underlying file
120 * system in vcache_reclaim().
121 * LOADED -> BLOCKED
122 * Either vcache_rekey*() is changing the vnode key or
123 * vrelel() is about to call VOP_INACTIVE().
124 * BLOCKED -> LOADED
125 * The block condition is over.
126 * LOADING -> RECLAIMED
127 * Either vcache_get() or vcache_new() failed to
128 * associate the underlying file system or vcache_rekey*()
129 * drops a vnode used as placeholder.
130 *
131 * Of these states LOADING, BLOCKED and RECLAIMING are intermediate
132 * and it is possible to wait for state change.
133 *
134 * State is protected with v_interlock with one exception:
135 * to change from LOADING both v_interlock and vcache_lock must be held
136 * so it is possible to check "state == LOADING" without holding
137 * v_interlock. See vcache_get() for details.
138 *
139 * Reference counting
140 *
141 * Vnode is considered active, if reference count (vnode_t::v_usecount)
142 * is non-zero. It is maintained using: vref(9) and vrele(9), as well
143 * as vput(9), routines. Common points holding references are e.g.
144 * file openings, current working directory, mount points, etc.
145 *
146 * v_usecount is adjusted with atomic operations, however to change
147 * from a non-zero value to zero the interlock must also be held.
148 */
149
150 #include <sys/cdefs.h>
151 __KERNEL_RCSID(0, "$NetBSD: vfs_vnode.c,v 1.147 2022/10/26 23:40:08 riastradh Exp $");
152
153 #ifdef _KERNEL_OPT
154 #include "opt_pax.h"
155 #endif
156
157 #include <sys/param.h>
158 #include <sys/kernel.h>
159
160 #include <sys/atomic.h>
161 #include <sys/buf.h>
162 #include <sys/conf.h>
163 #include <sys/device.h>
164 #include <sys/hash.h>
165 #include <sys/kauth.h>
166 #include <sys/kmem.h>
167 #include <sys/kthread.h>
168 #include <sys/module.h>
169 #include <sys/mount.h>
170 #include <sys/namei.h>
171 #include <sys/pax.h>
172 #include <sys/syscallargs.h>
173 #include <sys/sysctl.h>
174 #include <sys/systm.h>
175 #include <sys/vnode_impl.h>
176 #include <sys/wapbl.h>
177 #include <sys/fstrans.h>
178
179 #include <miscfs/deadfs/deadfs.h>
180 #include <miscfs/specfs/specdev.h>
181
182 #include <uvm/uvm.h>
183 #include <uvm/uvm_readahead.h>
184 #include <uvm/uvm_stat.h>
185
186 /* Flags to vrelel. */
187 #define VRELEL_ASYNC 0x0001 /* Always defer to vrele thread. */
188
189 #define LRU_VRELE 0
190 #define LRU_FREE 1
191 #define LRU_HOLD 2
192 #define LRU_COUNT 3
193
194 /*
195 * There are three lru lists: one holds vnodes waiting for async release,
196 * one is for vnodes which have no buffer/page references and one for those
197 * which do (i.e. v_holdcnt is non-zero). We put the lists into a single,
198 * private cache line as vnodes migrate between them while under the same
199 * lock (vdrain_lock).
200 */
201 u_int numvnodes __cacheline_aligned;
202 static vnodelst_t lru_list[LRU_COUNT] __cacheline_aligned;
203 static kmutex_t vdrain_lock __cacheline_aligned;
204 static kcondvar_t vdrain_cv;
205 static int vdrain_gen;
206 static kcondvar_t vdrain_gen_cv;
207 static bool vdrain_retry;
208 static lwp_t * vdrain_lwp;
209 SLIST_HEAD(hashhead, vnode_impl);
210 static kmutex_t vcache_lock __cacheline_aligned;
211 static kcondvar_t vcache_cv;
212 static u_int vcache_hashsize;
213 static u_long vcache_hashmask;
214 static struct hashhead *vcache_hashtab;
215 static pool_cache_t vcache_pool;
216 static void lru_requeue(vnode_t *, vnodelst_t *);
217 static vnodelst_t * lru_which(vnode_t *);
218 static vnode_impl_t * vcache_alloc(void);
219 static void vcache_dealloc(vnode_impl_t *);
220 static void vcache_free(vnode_impl_t *);
221 static void vcache_init(void);
222 static void vcache_reinit(void);
223 static void vcache_reclaim(vnode_t *);
224 static void vrelel(vnode_t *, int, int);
225 static void vdrain_thread(void *);
226 static void vnpanic(vnode_t *, const char *, ...)
227 __printflike(2, 3);
228
229 /* Routines having to do with the management of the vnode table. */
230
231 /*
232 * The high bit of v_usecount is a gate for vcache_tryvget(). It's set
233 * only when the vnode state is LOADED.
234 * The next bit of v_usecount is a flag for vrelel(). It's set
235 * from vcache_vget() and vcache_tryvget() whenever the operation succeeds.
236 */
237 #define VUSECOUNT_MASK 0x3fffffff
238 #define VUSECOUNT_GATE 0x80000000
239 #define VUSECOUNT_VGET 0x40000000
240
241 /*
242 * Return the current usecount of a vnode.
243 */
244 inline int
245 vrefcnt(struct vnode *vp)
246 {
247
248 return atomic_load_relaxed(&vp->v_usecount) & VUSECOUNT_MASK;
249 }
250
251 /* Vnode state operations and diagnostics. */
252
253 #if defined(DIAGNOSTIC)
254
255 #define VSTATE_VALID(state) \
256 ((state) != VS_ACTIVE && (state) != VS_MARKER)
257 #define VSTATE_GET(vp) \
258 vstate_assert_get((vp), __func__, __LINE__)
259 #define VSTATE_CHANGE(vp, from, to) \
260 vstate_assert_change((vp), (from), (to), __func__, __LINE__)
261 #define VSTATE_WAIT_STABLE(vp) \
262 vstate_assert_wait_stable((vp), __func__, __LINE__)
263
264 void
265 _vstate_assert(vnode_t *vp, enum vnode_state state, const char *func, int line,
266 bool has_lock)
267 {
268 vnode_impl_t *vip = VNODE_TO_VIMPL(vp);
269 int refcnt = vrefcnt(vp);
270
271 if (!has_lock) {
272 /*
273 * Prevent predictive loads from the CPU, but check the state
274 * without loooking first.
275 *
276 * XXX what does this pair with?
277 */
278 membar_enter();
279 if (state == VS_ACTIVE && refcnt > 0 &&
280 (vip->vi_state == VS_LOADED || vip->vi_state == VS_BLOCKED))
281 return;
282 if (vip->vi_state == state)
283 return;
284 mutex_enter((vp)->v_interlock);
285 }
286
287 KASSERTMSG(mutex_owned(vp->v_interlock), "at %s:%d", func, line);
288
289 if ((state == VS_ACTIVE && refcnt > 0 &&
290 (vip->vi_state == VS_LOADED || vip->vi_state == VS_BLOCKED)) ||
291 vip->vi_state == state) {
292 if (!has_lock)
293 mutex_exit((vp)->v_interlock);
294 return;
295 }
296 vnpanic(vp, "state is %s, usecount %d, expected %s at %s:%d",
297 vstate_name(vip->vi_state), refcnt,
298 vstate_name(state), func, line);
299 }
300
301 static enum vnode_state
302 vstate_assert_get(vnode_t *vp, const char *func, int line)
303 {
304 vnode_impl_t *vip = VNODE_TO_VIMPL(vp);
305
306 KASSERTMSG(mutex_owned(vp->v_interlock), "at %s:%d", func, line);
307 if (! VSTATE_VALID(vip->vi_state))
308 vnpanic(vp, "state is %s at %s:%d",
309 vstate_name(vip->vi_state), func, line);
310
311 return vip->vi_state;
312 }
313
314 static void
315 vstate_assert_wait_stable(vnode_t *vp, const char *func, int line)
316 {
317 vnode_impl_t *vip = VNODE_TO_VIMPL(vp);
318
319 KASSERTMSG(mutex_owned(vp->v_interlock), "at %s:%d", func, line);
320 if (! VSTATE_VALID(vip->vi_state))
321 vnpanic(vp, "state is %s at %s:%d",
322 vstate_name(vip->vi_state), func, line);
323
324 while (vip->vi_state != VS_LOADED && vip->vi_state != VS_RECLAIMED)
325 cv_wait(&vp->v_cv, vp->v_interlock);
326
327 if (! VSTATE_VALID(vip->vi_state))
328 vnpanic(vp, "state is %s at %s:%d",
329 vstate_name(vip->vi_state), func, line);
330 }
331
332 static void
333 vstate_assert_change(vnode_t *vp, enum vnode_state from, enum vnode_state to,
334 const char *func, int line)
335 {
336 bool gated = (atomic_load_relaxed(&vp->v_usecount) & VUSECOUNT_GATE);
337 vnode_impl_t *vip = VNODE_TO_VIMPL(vp);
338
339 KASSERTMSG(mutex_owned(vp->v_interlock), "at %s:%d", func, line);
340 if (from == VS_LOADING)
341 KASSERTMSG(mutex_owned(&vcache_lock), "at %s:%d", func, line);
342
343 if (! VSTATE_VALID(from))
344 vnpanic(vp, "from is %s at %s:%d",
345 vstate_name(from), func, line);
346 if (! VSTATE_VALID(to))
347 vnpanic(vp, "to is %s at %s:%d",
348 vstate_name(to), func, line);
349 if (vip->vi_state != from)
350 vnpanic(vp, "from is %s, expected %s at %s:%d\n",
351 vstate_name(vip->vi_state), vstate_name(from), func, line);
352 if ((from == VS_LOADED) != gated)
353 vnpanic(vp, "state is %s, gate %d does not match at %s:%d\n",
354 vstate_name(vip->vi_state), gated, func, line);
355
356 /* Open/close the gate for vcache_tryvget(). */
357 if (to == VS_LOADED) {
358 #ifndef __HAVE_ATOMIC_AS_MEMBAR
359 membar_release();
360 #endif
361 atomic_or_uint(&vp->v_usecount, VUSECOUNT_GATE);
362 } else {
363 atomic_and_uint(&vp->v_usecount, ~VUSECOUNT_GATE);
364 }
365
366 vip->vi_state = to;
367 if (from == VS_LOADING)
368 cv_broadcast(&vcache_cv);
369 if (to == VS_LOADED || to == VS_RECLAIMED)
370 cv_broadcast(&vp->v_cv);
371 }
372
373 #else /* defined(DIAGNOSTIC) */
374
375 #define VSTATE_GET(vp) \
376 (VNODE_TO_VIMPL((vp))->vi_state)
377 #define VSTATE_CHANGE(vp, from, to) \
378 vstate_change((vp), (from), (to))
379 #define VSTATE_WAIT_STABLE(vp) \
380 vstate_wait_stable((vp))
381 void
382 _vstate_assert(vnode_t *vp, enum vnode_state state, const char *func, int line,
383 bool has_lock)
384 {
385
386 }
387
388 static void
389 vstate_wait_stable(vnode_t *vp)
390 {
391 vnode_impl_t *vip = VNODE_TO_VIMPL(vp);
392
393 while (vip->vi_state != VS_LOADED && vip->vi_state != VS_RECLAIMED)
394 cv_wait(&vp->v_cv, vp->v_interlock);
395 }
396
397 static void
398 vstate_change(vnode_t *vp, enum vnode_state from, enum vnode_state to)
399 {
400 vnode_impl_t *vip = VNODE_TO_VIMPL(vp);
401
402 /* Open/close the gate for vcache_tryvget(). */
403 if (to == VS_LOADED) {
404 #ifndef __HAVE_ATOMIC_AS_MEMBAR
405 membar_release();
406 #endif
407 atomic_or_uint(&vp->v_usecount, VUSECOUNT_GATE);
408 } else {
409 atomic_and_uint(&vp->v_usecount, ~VUSECOUNT_GATE);
410 }
411
412 vip->vi_state = to;
413 if (from == VS_LOADING)
414 cv_broadcast(&vcache_cv);
415 if (to == VS_LOADED || to == VS_RECLAIMED)
416 cv_broadcast(&vp->v_cv);
417 }
418
419 #endif /* defined(DIAGNOSTIC) */
420
421 void
422 vfs_vnode_sysinit(void)
423 {
424 int error __diagused, i;
425
426 dead_rootmount = vfs_mountalloc(&dead_vfsops, NULL);
427 KASSERT(dead_rootmount != NULL);
428 dead_rootmount->mnt_iflag |= IMNT_MPSAFE;
429
430 mutex_init(&vdrain_lock, MUTEX_DEFAULT, IPL_NONE);
431 for (i = 0; i < LRU_COUNT; i++) {
432 TAILQ_INIT(&lru_list[i]);
433 }
434 vcache_init();
435
436 cv_init(&vdrain_cv, "vdrain");
437 cv_init(&vdrain_gen_cv, "vdrainwt");
438 error = kthread_create(PRI_VM, KTHREAD_MPSAFE, NULL, vdrain_thread,
439 NULL, &vdrain_lwp, "vdrain");
440 KASSERTMSG((error == 0), "kthread_create(vdrain) failed: %d", error);
441 }
442
443 /*
444 * Allocate a new marker vnode.
445 */
446 vnode_t *
447 vnalloc_marker(struct mount *mp)
448 {
449 vnode_impl_t *vip;
450 vnode_t *vp;
451
452 vip = pool_cache_get(vcache_pool, PR_WAITOK);
453 memset(vip, 0, sizeof(*vip));
454 vp = VIMPL_TO_VNODE(vip);
455 uvm_obj_init(&vp->v_uobj, &uvm_vnodeops, true, 1);
456 vp->v_mount = mp;
457 vp->v_type = VBAD;
458 vp->v_interlock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_NONE);
459 klist_init(&vip->vi_klist.vk_klist);
460 vp->v_klist = &vip->vi_klist;
461 vip->vi_state = VS_MARKER;
462
463 return vp;
464 }
465
466 /*
467 * Free a marker vnode.
468 */
469 void
470 vnfree_marker(vnode_t *vp)
471 {
472 vnode_impl_t *vip;
473
474 vip = VNODE_TO_VIMPL(vp);
475 KASSERT(vip->vi_state == VS_MARKER);
476 mutex_obj_free(vp->v_interlock);
477 uvm_obj_destroy(&vp->v_uobj, true);
478 klist_fini(&vip->vi_klist.vk_klist);
479 pool_cache_put(vcache_pool, vip);
480 }
481
482 /*
483 * Test a vnode for being a marker vnode.
484 */
485 bool
486 vnis_marker(vnode_t *vp)
487 {
488
489 return (VNODE_TO_VIMPL(vp)->vi_state == VS_MARKER);
490 }
491
492 /*
493 * Return the lru list this node should be on.
494 */
495 static vnodelst_t *
496 lru_which(vnode_t *vp)
497 {
498
499 KASSERT(mutex_owned(vp->v_interlock));
500
501 if (vp->v_holdcnt > 0)
502 return &lru_list[LRU_HOLD];
503 else
504 return &lru_list[LRU_FREE];
505 }
506
507 /*
508 * Put vnode to end of given list.
509 * Both the current and the new list may be NULL, used on vnode alloc/free.
510 * Adjust numvnodes and signal vdrain thread if there is work.
511 */
512 static void
513 lru_requeue(vnode_t *vp, vnodelst_t *listhd)
514 {
515 vnode_impl_t *vip;
516 int d;
517
518 /*
519 * If the vnode is on the correct list, and was put there recently,
520 * then leave it be, thus avoiding huge cache and lock contention.
521 */
522 vip = VNODE_TO_VIMPL(vp);
523 if (listhd == vip->vi_lrulisthd &&
524 (getticks() - vip->vi_lrulisttm) < hz) {
525 return;
526 }
527
528 mutex_enter(&vdrain_lock);
529 d = 0;
530 if (vip->vi_lrulisthd != NULL)
531 TAILQ_REMOVE(vip->vi_lrulisthd, vip, vi_lrulist);
532 else
533 d++;
534 vip->vi_lrulisthd = listhd;
535 vip->vi_lrulisttm = getticks();
536 if (vip->vi_lrulisthd != NULL)
537 TAILQ_INSERT_TAIL(vip->vi_lrulisthd, vip, vi_lrulist);
538 else
539 d--;
540 if (d != 0) {
541 /*
542 * Looks strange? This is not a bug. Don't store
543 * numvnodes unless there is a change - avoid false
544 * sharing on MP.
545 */
546 numvnodes += d;
547 }
548 if ((d > 0 && numvnodes > desiredvnodes) ||
549 listhd == &lru_list[LRU_VRELE])
550 cv_signal(&vdrain_cv);
551 mutex_exit(&vdrain_lock);
552 }
553
554 /*
555 * Release deferred vrele vnodes for this mount.
556 * Called with file system suspended.
557 */
558 void
559 vrele_flush(struct mount *mp)
560 {
561 vnode_impl_t *vip, *marker;
562 vnode_t *vp;
563 int when = 0;
564
565 KASSERT(fstrans_is_owner(mp));
566
567 marker = VNODE_TO_VIMPL(vnalloc_marker(NULL));
568
569 mutex_enter(&vdrain_lock);
570 TAILQ_INSERT_HEAD(&lru_list[LRU_VRELE], marker, vi_lrulist);
571
572 while ((vip = TAILQ_NEXT(marker, vi_lrulist))) {
573 TAILQ_REMOVE(&lru_list[LRU_VRELE], marker, vi_lrulist);
574 TAILQ_INSERT_AFTER(&lru_list[LRU_VRELE], vip, marker,
575 vi_lrulist);
576 vp = VIMPL_TO_VNODE(vip);
577 if (vnis_marker(vp))
578 continue;
579
580 KASSERT(vip->vi_lrulisthd == &lru_list[LRU_VRELE]);
581 TAILQ_REMOVE(vip->vi_lrulisthd, vip, vi_lrulist);
582 vip->vi_lrulisthd = &lru_list[LRU_HOLD];
583 vip->vi_lrulisttm = getticks();
584 TAILQ_INSERT_TAIL(vip->vi_lrulisthd, vip, vi_lrulist);
585 mutex_exit(&vdrain_lock);
586
587 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
588 mutex_enter(vp->v_interlock);
589 vrelel(vp, 0, LK_EXCLUSIVE);
590
591 if (getticks() > when) {
592 yield();
593 when = getticks() + hz / 10;
594 }
595
596 mutex_enter(&vdrain_lock);
597 }
598
599 TAILQ_REMOVE(&lru_list[LRU_VRELE], marker, vi_lrulist);
600 mutex_exit(&vdrain_lock);
601
602 vnfree_marker(VIMPL_TO_VNODE(marker));
603 }
604
605 /*
606 * Reclaim a cached vnode. Used from vdrain_thread only.
607 */
608 static __inline void
609 vdrain_remove(vnode_t *vp)
610 {
611 struct mount *mp;
612
613 KASSERT(mutex_owned(&vdrain_lock));
614
615 /* Probe usecount (unlocked). */
616 if (vrefcnt(vp) > 0)
617 return;
618 /* Try v_interlock -- we lock the wrong direction! */
619 if (!mutex_tryenter(vp->v_interlock))
620 return;
621 /* Probe usecount and state. */
622 if (vrefcnt(vp) > 0 || VSTATE_GET(vp) != VS_LOADED) {
623 mutex_exit(vp->v_interlock);
624 return;
625 }
626 mp = vp->v_mount;
627 if (fstrans_start_nowait(mp) != 0) {
628 mutex_exit(vp->v_interlock);
629 return;
630 }
631 vdrain_retry = true;
632 mutex_exit(&vdrain_lock);
633
634 if (vcache_vget(vp) == 0) {
635 if (!vrecycle(vp)) {
636 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
637 mutex_enter(vp->v_interlock);
638 vrelel(vp, 0, LK_EXCLUSIVE);
639 }
640 }
641 fstrans_done(mp);
642
643 mutex_enter(&vdrain_lock);
644 }
645
646 /*
647 * Release a cached vnode. Used from vdrain_thread only.
648 */
649 static __inline void
650 vdrain_vrele(vnode_t *vp)
651 {
652 vnode_impl_t *vip = VNODE_TO_VIMPL(vp);
653 struct mount *mp;
654
655 KASSERT(mutex_owned(&vdrain_lock));
656
657 mp = vp->v_mount;
658 if (fstrans_start_nowait(mp) != 0)
659 return;
660
661 /*
662 * First remove the vnode from the vrele list.
663 * Put it on the last lru list, the last vrele()
664 * will put it back onto the right list before
665 * its usecount reaches zero.
666 */
667 KASSERT(vip->vi_lrulisthd == &lru_list[LRU_VRELE]);
668 TAILQ_REMOVE(vip->vi_lrulisthd, vip, vi_lrulist);
669 vip->vi_lrulisthd = &lru_list[LRU_HOLD];
670 vip->vi_lrulisttm = getticks();
671 TAILQ_INSERT_TAIL(vip->vi_lrulisthd, vip, vi_lrulist);
672
673 vdrain_retry = true;
674 mutex_exit(&vdrain_lock);
675
676 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
677 mutex_enter(vp->v_interlock);
678 vrelel(vp, 0, LK_EXCLUSIVE);
679 fstrans_done(mp);
680
681 mutex_enter(&vdrain_lock);
682 }
683
684 /*
685 * Helper thread to keep the number of vnodes below desiredvnodes
686 * and release vnodes from asynchronous vrele.
687 */
688 static void
689 vdrain_thread(void *cookie)
690 {
691 int i;
692 u_int target;
693 vnode_impl_t *vip, *marker;
694
695 marker = VNODE_TO_VIMPL(vnalloc_marker(NULL));
696
697 mutex_enter(&vdrain_lock);
698
699 for (;;) {
700 vdrain_retry = false;
701 target = desiredvnodes - desiredvnodes/10;
702
703 for (i = 0; i < LRU_COUNT; i++) {
704 TAILQ_INSERT_HEAD(&lru_list[i], marker, vi_lrulist);
705 while ((vip = TAILQ_NEXT(marker, vi_lrulist))) {
706 TAILQ_REMOVE(&lru_list[i], marker, vi_lrulist);
707 TAILQ_INSERT_AFTER(&lru_list[i], vip, marker,
708 vi_lrulist);
709 if (vnis_marker(VIMPL_TO_VNODE(vip)))
710 continue;
711 if (i == LRU_VRELE)
712 vdrain_vrele(VIMPL_TO_VNODE(vip));
713 else if (numvnodes < target)
714 break;
715 else
716 vdrain_remove(VIMPL_TO_VNODE(vip));
717 }
718 TAILQ_REMOVE(&lru_list[i], marker, vi_lrulist);
719 }
720
721 if (vdrain_retry) {
722 kpause("vdrainrt", false, 1, &vdrain_lock);
723 } else {
724 vdrain_gen++;
725 cv_broadcast(&vdrain_gen_cv);
726 cv_wait(&vdrain_cv, &vdrain_lock);
727 }
728 }
729 }
730
731 /*
732 * Try to drop reference on a vnode. Abort if we are releasing the
733 * last reference. Note: this _must_ succeed if not the last reference.
734 */
735 static bool
736 vtryrele(vnode_t *vp)
737 {
738 u_int use, next;
739
740 #ifndef __HAVE_ATOMIC_AS_MEMBAR
741 membar_release();
742 #endif
743 for (use = atomic_load_relaxed(&vp->v_usecount);; use = next) {
744 if (__predict_false((use & VUSECOUNT_MASK) == 1)) {
745 return false;
746 }
747 KASSERT((use & VUSECOUNT_MASK) > 1);
748 next = atomic_cas_uint(&vp->v_usecount, use, use - 1);
749 if (__predict_true(next == use)) {
750 return true;
751 }
752 }
753 }
754
755 /*
756 * vput: unlock and release the reference.
757 */
758 void
759 vput(vnode_t *vp)
760 {
761 int lktype;
762
763 /*
764 * Do an unlocked check of the usecount. If it looks like we're not
765 * about to drop the last reference, then unlock the vnode and try
766 * to drop the reference. If it ends up being the last reference
767 * after all, vrelel() can fix it all up. Most of the time this
768 * will all go to plan.
769 */
770 if (vrefcnt(vp) > 1) {
771 VOP_UNLOCK(vp);
772 if (vtryrele(vp)) {
773 return;
774 }
775 lktype = LK_NONE;
776 } else {
777 lktype = VOP_ISLOCKED(vp);
778 KASSERT(lktype != LK_NONE);
779 }
780 mutex_enter(vp->v_interlock);
781 vrelel(vp, 0, lktype);
782 }
783
784 /*
785 * Vnode release. If reference count drops to zero, call inactive
786 * routine and either return to freelist or free to the pool.
787 */
788 static void
789 vrelel(vnode_t *vp, int flags, int lktype)
790 {
791 const bool async = ((flags & VRELEL_ASYNC) != 0);
792 bool recycle, defer, objlock_held;
793 u_int use, next;
794 int error;
795
796 objlock_held = false;
797
798 retry:
799 KASSERT(mutex_owned(vp->v_interlock));
800
801 if (__predict_false(vp->v_op == dead_vnodeop_p &&
802 VSTATE_GET(vp) != VS_RECLAIMED)) {
803 vnpanic(vp, "dead but not clean");
804 }
805
806 /*
807 * If not the last reference, just unlock and drop the reference count.
808 *
809 * Otherwise make sure we pass a point in time where we hold the
810 * last reference with VGET flag unset.
811 */
812 for (use = atomic_load_relaxed(&vp->v_usecount);; use = next) {
813 if (__predict_false((use & VUSECOUNT_MASK) > 1)) {
814 if (objlock_held) {
815 objlock_held = false;
816 rw_exit(vp->v_uobj.vmobjlock);
817 }
818 if (lktype != LK_NONE) {
819 mutex_exit(vp->v_interlock);
820 lktype = LK_NONE;
821 VOP_UNLOCK(vp);
822 mutex_enter(vp->v_interlock);
823 }
824 if (vtryrele(vp)) {
825 mutex_exit(vp->v_interlock);
826 return;
827 }
828 next = atomic_load_relaxed(&vp->v_usecount);
829 continue;
830 }
831 KASSERT((use & VUSECOUNT_MASK) == 1);
832 next = use & ~VUSECOUNT_VGET;
833 if (next != use) {
834 next = atomic_cas_uint(&vp->v_usecount, use, next);
835 }
836 if (__predict_true(next == use)) {
837 break;
838 }
839 }
840 #ifndef __HAVE_ATOMIC_AS_MEMBAR
841 membar_acquire();
842 #endif
843 if (vrefcnt(vp) <= 0 || vp->v_writecount != 0) {
844 vnpanic(vp, "%s: bad ref count", __func__);
845 }
846
847 #ifdef DIAGNOSTIC
848 if ((vp->v_type == VBLK || vp->v_type == VCHR) &&
849 vp->v_specnode != NULL && vp->v_specnode->sn_opencnt != 0) {
850 vprint("vrelel: missing VOP_CLOSE()", vp);
851 }
852 #endif
853
854 /*
855 * If already clean there is no need to lock, defer or
856 * deactivate this node.
857 */
858 if (VSTATE_GET(vp) == VS_RECLAIMED) {
859 if (objlock_held) {
860 objlock_held = false;
861 rw_exit(vp->v_uobj.vmobjlock);
862 }
863 if (lktype != LK_NONE) {
864 mutex_exit(vp->v_interlock);
865 lktype = LK_NONE;
866 VOP_UNLOCK(vp);
867 mutex_enter(vp->v_interlock);
868 }
869 goto out;
870 }
871
872 /*
873 * First try to get the vnode locked for VOP_INACTIVE().
874 * Defer vnode release to vdrain_thread if caller requests
875 * it explicitly, is the pagedaemon or the lock failed.
876 */
877 defer = false;
878 if ((curlwp == uvm.pagedaemon_lwp) || async) {
879 defer = true;
880 } else if (lktype == LK_SHARED) {
881 /* Excellent chance of getting, if the last ref. */
882 error = vn_lock(vp, LK_UPGRADE | LK_RETRY | LK_NOWAIT);
883 if (error != 0) {
884 defer = true;
885 } else {
886 lktype = LK_EXCLUSIVE;
887 }
888 } else if (lktype == LK_NONE) {
889 /* Excellent chance of getting, if the last ref. */
890 error = vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_NOWAIT);
891 if (error != 0) {
892 defer = true;
893 } else {
894 lktype = LK_EXCLUSIVE;
895 }
896 }
897 KASSERT(mutex_owned(vp->v_interlock));
898 if (defer) {
899 /*
900 * Defer reclaim to the kthread; it's not safe to
901 * clean it here. We donate it our last reference.
902 */
903 if (lktype != LK_NONE) {
904 mutex_exit(vp->v_interlock);
905 VOP_UNLOCK(vp);
906 mutex_enter(vp->v_interlock);
907 }
908 lru_requeue(vp, &lru_list[LRU_VRELE]);
909 mutex_exit(vp->v_interlock);
910 return;
911 }
912 KASSERT(lktype == LK_EXCLUSIVE);
913
914 /* If the node gained another reference, retry. */
915 use = atomic_load_relaxed(&vp->v_usecount);
916 if ((use & VUSECOUNT_VGET) != 0) {
917 goto retry;
918 }
919 KASSERT((use & VUSECOUNT_MASK) == 1);
920
921 if ((vp->v_iflag & (VI_TEXT|VI_EXECMAP|VI_WRMAP)) != 0 ||
922 (vp->v_vflag & VV_MAPPED) != 0) {
923 /* Take care of space accounting. */
924 if (!objlock_held) {
925 objlock_held = true;
926 if (!rw_tryenter(vp->v_uobj.vmobjlock, RW_WRITER)) {
927 mutex_exit(vp->v_interlock);
928 rw_enter(vp->v_uobj.vmobjlock, RW_WRITER);
929 mutex_enter(vp->v_interlock);
930 goto retry;
931 }
932 }
933 if ((vp->v_iflag & VI_EXECMAP) != 0) {
934 cpu_count(CPU_COUNT_EXECPAGES, -vp->v_uobj.uo_npages);
935 }
936 vp->v_iflag &= ~(VI_TEXT|VI_EXECMAP|VI_WRMAP);
937 vp->v_vflag &= ~VV_MAPPED;
938 }
939 if (objlock_held) {
940 objlock_held = false;
941 rw_exit(vp->v_uobj.vmobjlock);
942 }
943
944 /*
945 * Deactivate the vnode, but preserve our reference across
946 * the call to VOP_INACTIVE().
947 *
948 * If VOP_INACTIVE() indicates that the file has been
949 * deleted, then recycle the vnode.
950 *
951 * Note that VOP_INACTIVE() will not drop the vnode lock.
952 */
953 mutex_exit(vp->v_interlock);
954 recycle = false;
955 VOP_INACTIVE(vp, &recycle);
956 if (!recycle) {
957 lktype = LK_NONE;
958 VOP_UNLOCK(vp);
959 }
960 mutex_enter(vp->v_interlock);
961
962 /*
963 * Block new references then check again to see if a
964 * new reference was acquired in the meantime. If
965 * it was, restore the vnode state and try again.
966 */
967 if (recycle) {
968 VSTATE_CHANGE(vp, VS_LOADED, VS_BLOCKED);
969 use = atomic_load_relaxed(&vp->v_usecount);
970 if ((use & VUSECOUNT_VGET) != 0) {
971 VSTATE_CHANGE(vp, VS_BLOCKED, VS_LOADED);
972 goto retry;
973 }
974 KASSERT((use & VUSECOUNT_MASK) == 1);
975 }
976
977 /*
978 * Recycle the vnode if the file is now unused (unlinked).
979 */
980 if (recycle) {
981 VSTATE_ASSERT(vp, VS_BLOCKED);
982 KASSERT(lktype == LK_EXCLUSIVE);
983 /* vcache_reclaim drops the lock. */
984 lktype = LK_NONE;
985 vcache_reclaim(vp);
986 }
987 KASSERT(vrefcnt(vp) > 0);
988 KASSERT(lktype == LK_NONE);
989
990 out:
991 for (use = atomic_load_relaxed(&vp->v_usecount);; use = next) {
992 if (__predict_false((use & VUSECOUNT_VGET) != 0 &&
993 (use & VUSECOUNT_MASK) == 1)) {
994 /* Gained and released another reference, retry. */
995 goto retry;
996 }
997 next = atomic_cas_uint(&vp->v_usecount, use, use - 1);
998 if (__predict_true(next == use)) {
999 if (__predict_false((use & VUSECOUNT_MASK) != 1)) {
1000 /* Gained another reference. */
1001 mutex_exit(vp->v_interlock);
1002 return;
1003 }
1004 break;
1005 }
1006 }
1007 #ifndef __HAVE_ATOMIC_AS_MEMBAR
1008 membar_acquire();
1009 #endif
1010
1011 if (VSTATE_GET(vp) == VS_RECLAIMED && vp->v_holdcnt == 0) {
1012 /*
1013 * It's clean so destroy it. It isn't referenced
1014 * anywhere since it has been reclaimed.
1015 */
1016 vcache_free(VNODE_TO_VIMPL(vp));
1017 } else {
1018 /*
1019 * Otherwise, put it back onto the freelist. It
1020 * can't be destroyed while still associated with
1021 * a file system.
1022 */
1023 lru_requeue(vp, lru_which(vp));
1024 mutex_exit(vp->v_interlock);
1025 }
1026 }
1027
1028 void
1029 vrele(vnode_t *vp)
1030 {
1031
1032 if (vtryrele(vp)) {
1033 return;
1034 }
1035 mutex_enter(vp->v_interlock);
1036 vrelel(vp, 0, LK_NONE);
1037 }
1038
1039 /*
1040 * Asynchronous vnode release, vnode is released in different context.
1041 */
1042 void
1043 vrele_async(vnode_t *vp)
1044 {
1045
1046 if (vtryrele(vp)) {
1047 return;
1048 }
1049 mutex_enter(vp->v_interlock);
1050 vrelel(vp, VRELEL_ASYNC, LK_NONE);
1051 }
1052
1053 /*
1054 * Vnode reference, where a reference is already held by some other
1055 * object (for example, a file structure).
1056 *
1057 * NB: lockless code sequences may rely on this not blocking.
1058 */
1059 void
1060 vref(vnode_t *vp)
1061 {
1062
1063 KASSERT(vrefcnt(vp) > 0);
1064
1065 atomic_inc_uint(&vp->v_usecount);
1066 }
1067
1068 /*
1069 * Page or buffer structure gets a reference.
1070 * Called with v_interlock held.
1071 */
1072 void
1073 vholdl(vnode_t *vp)
1074 {
1075
1076 KASSERT(mutex_owned(vp->v_interlock));
1077
1078 if (vp->v_holdcnt++ == 0 && vrefcnt(vp) == 0)
1079 lru_requeue(vp, lru_which(vp));
1080 }
1081
1082 /*
1083 * Page or buffer structure gets a reference.
1084 */
1085 void
1086 vhold(vnode_t *vp)
1087 {
1088
1089 mutex_enter(vp->v_interlock);
1090 vholdl(vp);
1091 mutex_exit(vp->v_interlock);
1092 }
1093
1094 /*
1095 * Page or buffer structure frees a reference.
1096 * Called with v_interlock held.
1097 */
1098 void
1099 holdrelel(vnode_t *vp)
1100 {
1101
1102 KASSERT(mutex_owned(vp->v_interlock));
1103
1104 if (vp->v_holdcnt <= 0) {
1105 vnpanic(vp, "%s: holdcnt vp %p", __func__, vp);
1106 }
1107
1108 vp->v_holdcnt--;
1109 if (vp->v_holdcnt == 0 && vrefcnt(vp) == 0)
1110 lru_requeue(vp, lru_which(vp));
1111 }
1112
1113 /*
1114 * Page or buffer structure frees a reference.
1115 */
1116 void
1117 holdrele(vnode_t *vp)
1118 {
1119
1120 mutex_enter(vp->v_interlock);
1121 holdrelel(vp);
1122 mutex_exit(vp->v_interlock);
1123 }
1124
1125 /*
1126 * Recycle an unused vnode if caller holds the last reference.
1127 */
1128 bool
1129 vrecycle(vnode_t *vp)
1130 {
1131 int error __diagused;
1132
1133 mutex_enter(vp->v_interlock);
1134
1135 /* If the vnode is already clean we're done. */
1136 VSTATE_WAIT_STABLE(vp);
1137 if (VSTATE_GET(vp) != VS_LOADED) {
1138 VSTATE_ASSERT(vp, VS_RECLAIMED);
1139 vrelel(vp, 0, LK_NONE);
1140 return true;
1141 }
1142
1143 /* Prevent further references until the vnode is locked. */
1144 VSTATE_CHANGE(vp, VS_LOADED, VS_BLOCKED);
1145
1146 /* Make sure we hold the last reference. */
1147 if (vrefcnt(vp) != 1) {
1148 VSTATE_CHANGE(vp, VS_BLOCKED, VS_LOADED);
1149 mutex_exit(vp->v_interlock);
1150 return false;
1151 }
1152
1153 mutex_exit(vp->v_interlock);
1154
1155 /*
1156 * On a leaf file system this lock will always succeed as we hold
1157 * the last reference and prevent further references.
1158 * On layered file systems waiting for the lock would open a can of
1159 * deadlocks as the lower vnodes may have other active references.
1160 */
1161 error = vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_NOWAIT);
1162
1163 mutex_enter(vp->v_interlock);
1164 if (error) {
1165 VSTATE_CHANGE(vp, VS_BLOCKED, VS_LOADED);
1166 mutex_exit(vp->v_interlock);
1167 return false;
1168 }
1169
1170 KASSERT(vrefcnt(vp) == 1);
1171 vcache_reclaim(vp);
1172 vrelel(vp, 0, LK_NONE);
1173
1174 return true;
1175 }
1176
1177 /*
1178 * Helper for vrevoke() to propagate suspension from lastmp
1179 * to thismp. Both args may be NULL.
1180 * Returns the currently suspended file system or NULL.
1181 */
1182 static struct mount *
1183 vrevoke_suspend_next(struct mount *lastmp, struct mount *thismp)
1184 {
1185 int error;
1186
1187 if (lastmp == thismp)
1188 return thismp;
1189
1190 if (lastmp != NULL)
1191 vfs_resume(lastmp);
1192
1193 if (thismp == NULL)
1194 return NULL;
1195
1196 do {
1197 error = vfs_suspend(thismp, 0);
1198 } while (error == EINTR || error == ERESTART);
1199
1200 if (error == 0)
1201 return thismp;
1202
1203 KASSERT(error == EOPNOTSUPP || error == ENOENT);
1204 return NULL;
1205 }
1206
1207 /*
1208 * Eliminate all activity associated with the requested vnode
1209 * and with all vnodes aliased to the requested vnode.
1210 */
1211 void
1212 vrevoke(vnode_t *vp)
1213 {
1214 struct mount *mp;
1215 vnode_t *vq;
1216 enum vtype type;
1217 dev_t dev;
1218
1219 KASSERT(vrefcnt(vp) > 0);
1220
1221 mp = vrevoke_suspend_next(NULL, vp->v_mount);
1222
1223 mutex_enter(vp->v_interlock);
1224 VSTATE_WAIT_STABLE(vp);
1225 if (VSTATE_GET(vp) == VS_RECLAIMED) {
1226 mutex_exit(vp->v_interlock);
1227 } else if (vp->v_type != VBLK && vp->v_type != VCHR) {
1228 atomic_inc_uint(&vp->v_usecount);
1229 mutex_exit(vp->v_interlock);
1230 vgone(vp);
1231 } else {
1232 dev = vp->v_rdev;
1233 type = vp->v_type;
1234 mutex_exit(vp->v_interlock);
1235
1236 while (spec_node_lookup_by_dev(type, dev, VDEAD_NOWAIT, &vq)
1237 == 0) {
1238 mp = vrevoke_suspend_next(mp, vq->v_mount);
1239 vgone(vq);
1240 }
1241 }
1242 vrevoke_suspend_next(mp, NULL);
1243 }
1244
1245 /*
1246 * Eliminate all activity associated with a vnode in preparation for
1247 * reuse. Drops a reference from the vnode.
1248 */
1249 void
1250 vgone(vnode_t *vp)
1251 {
1252 int lktype;
1253
1254 KASSERT(vp->v_mount == dead_rootmount || fstrans_is_owner(vp->v_mount));
1255
1256 vn_lock(vp, LK_EXCLUSIVE | LK_RETRY);
1257 lktype = LK_EXCLUSIVE;
1258 mutex_enter(vp->v_interlock);
1259 VSTATE_WAIT_STABLE(vp);
1260 if (VSTATE_GET(vp) == VS_LOADED) {
1261 VSTATE_CHANGE(vp, VS_LOADED, VS_BLOCKED);
1262 vcache_reclaim(vp);
1263 lktype = LK_NONE;
1264 }
1265 VSTATE_ASSERT(vp, VS_RECLAIMED);
1266 vrelel(vp, 0, lktype);
1267 }
1268
1269 static inline uint32_t
1270 vcache_hash(const struct vcache_key *key)
1271 {
1272 uint32_t hash = HASH32_BUF_INIT;
1273
1274 KASSERT(key->vk_key_len > 0);
1275
1276 hash = hash32_buf(&key->vk_mount, sizeof(struct mount *), hash);
1277 hash = hash32_buf(key->vk_key, key->vk_key_len, hash);
1278 return hash;
1279 }
1280
1281 static int
1282 vcache_stats(struct hashstat_sysctl *hs, bool fill)
1283 {
1284 vnode_impl_t *vip;
1285 uint64_t chain;
1286
1287 strlcpy(hs->hash_name, "vcache", sizeof(hs->hash_name));
1288 strlcpy(hs->hash_desc, "vnode cache hash", sizeof(hs->hash_desc));
1289 if (!fill)
1290 return 0;
1291
1292 hs->hash_size = vcache_hashmask + 1;
1293
1294 for (size_t i = 0; i < hs->hash_size; i++) {
1295 chain = 0;
1296 mutex_enter(&vcache_lock);
1297 SLIST_FOREACH(vip, &vcache_hashtab[i], vi_hash) {
1298 chain++;
1299 }
1300 mutex_exit(&vcache_lock);
1301 if (chain > 0) {
1302 hs->hash_used++;
1303 hs->hash_items += chain;
1304 if (chain > hs->hash_maxchain)
1305 hs->hash_maxchain = chain;
1306 }
1307 preempt_point();
1308 }
1309
1310 return 0;
1311 }
1312
1313 static void
1314 vcache_init(void)
1315 {
1316
1317 vcache_pool = pool_cache_init(sizeof(vnode_impl_t), coherency_unit,
1318 0, 0, "vcachepl", NULL, IPL_NONE, NULL, NULL, NULL);
1319 KASSERT(vcache_pool != NULL);
1320 mutex_init(&vcache_lock, MUTEX_DEFAULT, IPL_NONE);
1321 cv_init(&vcache_cv, "vcache");
1322 vcache_hashsize = desiredvnodes;
1323 vcache_hashtab = hashinit(desiredvnodes, HASH_SLIST, true,
1324 &vcache_hashmask);
1325 hashstat_register("vcache", vcache_stats);
1326 }
1327
1328 static void
1329 vcache_reinit(void)
1330 {
1331 int i;
1332 uint32_t hash;
1333 u_long oldmask, newmask;
1334 struct hashhead *oldtab, *newtab;
1335 vnode_impl_t *vip;
1336
1337 newtab = hashinit(desiredvnodes, HASH_SLIST, true, &newmask);
1338 mutex_enter(&vcache_lock);
1339 oldtab = vcache_hashtab;
1340 oldmask = vcache_hashmask;
1341 vcache_hashsize = desiredvnodes;
1342 vcache_hashtab = newtab;
1343 vcache_hashmask = newmask;
1344 for (i = 0; i <= oldmask; i++) {
1345 while ((vip = SLIST_FIRST(&oldtab[i])) != NULL) {
1346 SLIST_REMOVE(&oldtab[i], vip, vnode_impl, vi_hash);
1347 hash = vcache_hash(&vip->vi_key);
1348 SLIST_INSERT_HEAD(&newtab[hash & vcache_hashmask],
1349 vip, vi_hash);
1350 }
1351 }
1352 mutex_exit(&vcache_lock);
1353 hashdone(oldtab, HASH_SLIST, oldmask);
1354 }
1355
1356 static inline vnode_impl_t *
1357 vcache_hash_lookup(const struct vcache_key *key, uint32_t hash)
1358 {
1359 struct hashhead *hashp;
1360 vnode_impl_t *vip;
1361
1362 KASSERT(mutex_owned(&vcache_lock));
1363
1364 hashp = &vcache_hashtab[hash & vcache_hashmask];
1365 SLIST_FOREACH(vip, hashp, vi_hash) {
1366 if (key->vk_mount != vip->vi_key.vk_mount)
1367 continue;
1368 if (key->vk_key_len != vip->vi_key.vk_key_len)
1369 continue;
1370 if (memcmp(key->vk_key, vip->vi_key.vk_key, key->vk_key_len))
1371 continue;
1372 return vip;
1373 }
1374 return NULL;
1375 }
1376
1377 /*
1378 * Allocate a new, uninitialized vcache node.
1379 */
1380 static vnode_impl_t *
1381 vcache_alloc(void)
1382 {
1383 vnode_impl_t *vip;
1384 vnode_t *vp;
1385
1386 vip = pool_cache_get(vcache_pool, PR_WAITOK);
1387 vp = VIMPL_TO_VNODE(vip);
1388 memset(vip, 0, sizeof(*vip));
1389
1390 rw_init(&vip->vi_lock);
1391 vp->v_interlock = mutex_obj_alloc(MUTEX_DEFAULT, IPL_NONE);
1392
1393 uvm_obj_init(&vp->v_uobj, &uvm_vnodeops, true, 1);
1394 klist_init(&vip->vi_klist.vk_klist);
1395 vp->v_klist = &vip->vi_klist;
1396 cv_init(&vp->v_cv, "vnode");
1397 cache_vnode_init(vp);
1398
1399 vp->v_usecount = 1;
1400 vp->v_type = VNON;
1401 vp->v_size = vp->v_writesize = VSIZENOTSET;
1402
1403 vip->vi_state = VS_LOADING;
1404
1405 lru_requeue(vp, &lru_list[LRU_FREE]);
1406
1407 return vip;
1408 }
1409
1410 /*
1411 * Deallocate a vcache node in state VS_LOADING.
1412 *
1413 * vcache_lock held on entry and released on return.
1414 */
1415 static void
1416 vcache_dealloc(vnode_impl_t *vip)
1417 {
1418 vnode_t *vp;
1419
1420 KASSERT(mutex_owned(&vcache_lock));
1421
1422 vp = VIMPL_TO_VNODE(vip);
1423 vfs_ref(dead_rootmount);
1424 vfs_insmntque(vp, dead_rootmount);
1425 mutex_enter(vp->v_interlock);
1426 vp->v_op = dead_vnodeop_p;
1427 VSTATE_CHANGE(vp, VS_LOADING, VS_RECLAIMED);
1428 mutex_exit(&vcache_lock);
1429 vrelel(vp, 0, LK_NONE);
1430 }
1431
1432 /*
1433 * Free an unused, unreferenced vcache node.
1434 * v_interlock locked on entry.
1435 */
1436 static void
1437 vcache_free(vnode_impl_t *vip)
1438 {
1439 vnode_t *vp;
1440
1441 vp = VIMPL_TO_VNODE(vip);
1442 KASSERT(mutex_owned(vp->v_interlock));
1443
1444 KASSERT(vrefcnt(vp) == 0);
1445 KASSERT(vp->v_holdcnt == 0);
1446 KASSERT(vp->v_writecount == 0);
1447 lru_requeue(vp, NULL);
1448 mutex_exit(vp->v_interlock);
1449
1450 vfs_insmntque(vp, NULL);
1451 if (vp->v_type == VBLK || vp->v_type == VCHR)
1452 spec_node_destroy(vp);
1453
1454 mutex_obj_free(vp->v_interlock);
1455 rw_destroy(&vip->vi_lock);
1456 uvm_obj_destroy(&vp->v_uobj, true);
1457 KASSERT(vp->v_klist == &vip->vi_klist);
1458 klist_fini(&vip->vi_klist.vk_klist);
1459 cv_destroy(&vp->v_cv);
1460 cache_vnode_fini(vp);
1461 pool_cache_put(vcache_pool, vip);
1462 }
1463
1464 /*
1465 * Try to get an initial reference on this cached vnode.
1466 * Returns zero on success or EBUSY if the vnode state is not LOADED.
1467 *
1468 * NB: lockless code sequences may rely on this not blocking.
1469 */
1470 int
1471 vcache_tryvget(vnode_t *vp)
1472 {
1473 u_int use, next;
1474
1475 for (use = atomic_load_relaxed(&vp->v_usecount);; use = next) {
1476 if (__predict_false((use & VUSECOUNT_GATE) == 0)) {
1477 return EBUSY;
1478 }
1479 next = atomic_cas_uint(&vp->v_usecount,
1480 use, (use + 1) | VUSECOUNT_VGET);
1481 if (__predict_true(next == use)) {
1482 #ifndef __HAVE_ATOMIC_AS_MEMBAR
1483 membar_acquire();
1484 #endif
1485 return 0;
1486 }
1487 }
1488 }
1489
1490 /*
1491 * Try to get an initial reference on this cached vnode.
1492 * Returns zero on success and ENOENT if the vnode has been reclaimed.
1493 * Will wait for the vnode state to be stable.
1494 *
1495 * v_interlock locked on entry and unlocked on exit.
1496 */
1497 int
1498 vcache_vget(vnode_t *vp)
1499 {
1500 int error;
1501
1502 KASSERT(mutex_owned(vp->v_interlock));
1503
1504 /* Increment hold count to prevent vnode from disappearing. */
1505 vp->v_holdcnt++;
1506 VSTATE_WAIT_STABLE(vp);
1507 vp->v_holdcnt--;
1508
1509 /* If this was the last reference to a reclaimed vnode free it now. */
1510 if (__predict_false(VSTATE_GET(vp) == VS_RECLAIMED)) {
1511 if (vp->v_holdcnt == 0 && vrefcnt(vp) == 0)
1512 vcache_free(VNODE_TO_VIMPL(vp));
1513 else
1514 mutex_exit(vp->v_interlock);
1515 return ENOENT;
1516 }
1517 VSTATE_ASSERT(vp, VS_LOADED);
1518 error = vcache_tryvget(vp);
1519 KASSERT(error == 0);
1520 mutex_exit(vp->v_interlock);
1521
1522 return 0;
1523 }
1524
1525 /*
1526 * Get a vnode / fs node pair by key and return it referenced through vpp.
1527 */
1528 int
1529 vcache_get(struct mount *mp, const void *key, size_t key_len,
1530 struct vnode **vpp)
1531 {
1532 int error;
1533 uint32_t hash;
1534 const void *new_key;
1535 struct vnode *vp;
1536 struct vcache_key vcache_key;
1537 vnode_impl_t *vip, *new_vip;
1538
1539 new_key = NULL;
1540 *vpp = NULL;
1541
1542 vcache_key.vk_mount = mp;
1543 vcache_key.vk_key = key;
1544 vcache_key.vk_key_len = key_len;
1545 hash = vcache_hash(&vcache_key);
1546
1547 again:
1548 mutex_enter(&vcache_lock);
1549 vip = vcache_hash_lookup(&vcache_key, hash);
1550
1551 /* If found, take a reference or retry. */
1552 if (__predict_true(vip != NULL)) {
1553 /*
1554 * If the vnode is loading we cannot take the v_interlock
1555 * here as it might change during load (see uvm_obj_setlock()).
1556 * As changing state from VS_LOADING requires both vcache_lock
1557 * and v_interlock it is safe to test with vcache_lock held.
1558 *
1559 * Wait for vnodes changing state from VS_LOADING and retry.
1560 */
1561 if (__predict_false(vip->vi_state == VS_LOADING)) {
1562 cv_wait(&vcache_cv, &vcache_lock);
1563 mutex_exit(&vcache_lock);
1564 goto again;
1565 }
1566 vp = VIMPL_TO_VNODE(vip);
1567 mutex_enter(vp->v_interlock);
1568 mutex_exit(&vcache_lock);
1569 error = vcache_vget(vp);
1570 if (error == ENOENT)
1571 goto again;
1572 if (error == 0)
1573 *vpp = vp;
1574 KASSERT((error != 0) == (*vpp == NULL));
1575 return error;
1576 }
1577 mutex_exit(&vcache_lock);
1578
1579 /* Allocate and initialize a new vcache / vnode pair. */
1580 error = vfs_busy(mp);
1581 if (error)
1582 return error;
1583 new_vip = vcache_alloc();
1584 new_vip->vi_key = vcache_key;
1585 vp = VIMPL_TO_VNODE(new_vip);
1586 mutex_enter(&vcache_lock);
1587 vip = vcache_hash_lookup(&vcache_key, hash);
1588 if (vip == NULL) {
1589 SLIST_INSERT_HEAD(&vcache_hashtab[hash & vcache_hashmask],
1590 new_vip, vi_hash);
1591 vip = new_vip;
1592 }
1593
1594 /* If another thread beat us inserting this node, retry. */
1595 if (vip != new_vip) {
1596 vcache_dealloc(new_vip);
1597 vfs_unbusy(mp);
1598 goto again;
1599 }
1600 mutex_exit(&vcache_lock);
1601
1602 /* Load the fs node. Exclusive as new_node is VS_LOADING. */
1603 error = VFS_LOADVNODE(mp, vp, key, key_len, &new_key);
1604 if (error) {
1605 mutex_enter(&vcache_lock);
1606 SLIST_REMOVE(&vcache_hashtab[hash & vcache_hashmask],
1607 new_vip, vnode_impl, vi_hash);
1608 vcache_dealloc(new_vip);
1609 vfs_unbusy(mp);
1610 KASSERT(*vpp == NULL);
1611 return error;
1612 }
1613 KASSERT(new_key != NULL);
1614 KASSERT(memcmp(key, new_key, key_len) == 0);
1615 KASSERT(vp->v_op != NULL);
1616 vfs_insmntque(vp, mp);
1617 if ((mp->mnt_iflag & IMNT_MPSAFE) != 0)
1618 vp->v_vflag |= VV_MPSAFE;
1619 vfs_ref(mp);
1620 vfs_unbusy(mp);
1621
1622 /* Finished loading, finalize node. */
1623 mutex_enter(&vcache_lock);
1624 new_vip->vi_key.vk_key = new_key;
1625 mutex_enter(vp->v_interlock);
1626 VSTATE_CHANGE(vp, VS_LOADING, VS_LOADED);
1627 mutex_exit(vp->v_interlock);
1628 mutex_exit(&vcache_lock);
1629 *vpp = vp;
1630 return 0;
1631 }
1632
1633 /*
1634 * Create a new vnode / fs node pair and return it referenced through vpp.
1635 */
1636 int
1637 vcache_new(struct mount *mp, struct vnode *dvp, struct vattr *vap,
1638 kauth_cred_t cred, void *extra, struct vnode **vpp)
1639 {
1640 int error;
1641 uint32_t hash;
1642 struct vnode *vp, *ovp;
1643 vnode_impl_t *vip, *ovip;
1644
1645 *vpp = NULL;
1646
1647 /* Allocate and initialize a new vcache / vnode pair. */
1648 error = vfs_busy(mp);
1649 if (error)
1650 return error;
1651 vip = vcache_alloc();
1652 vip->vi_key.vk_mount = mp;
1653 vp = VIMPL_TO_VNODE(vip);
1654
1655 /* Create and load the fs node. */
1656 error = VFS_NEWVNODE(mp, dvp, vp, vap, cred, extra,
1657 &vip->vi_key.vk_key_len, &vip->vi_key.vk_key);
1658 if (error) {
1659 mutex_enter(&vcache_lock);
1660 vcache_dealloc(vip);
1661 vfs_unbusy(mp);
1662 KASSERT(*vpp == NULL);
1663 return error;
1664 }
1665 KASSERT(vp->v_op != NULL);
1666 KASSERT((vip->vi_key.vk_key_len == 0) == (mp == dead_rootmount));
1667 if (vip->vi_key.vk_key_len > 0) {
1668 KASSERT(vip->vi_key.vk_key != NULL);
1669 hash = vcache_hash(&vip->vi_key);
1670
1671 /*
1672 * Wait for previous instance to be reclaimed,
1673 * then insert new node.
1674 */
1675 mutex_enter(&vcache_lock);
1676 while ((ovip = vcache_hash_lookup(&vip->vi_key, hash))) {
1677 ovp = VIMPL_TO_VNODE(ovip);
1678 mutex_enter(ovp->v_interlock);
1679 mutex_exit(&vcache_lock);
1680 error = vcache_vget(ovp);
1681 KASSERT(error == ENOENT);
1682 mutex_enter(&vcache_lock);
1683 }
1684 SLIST_INSERT_HEAD(&vcache_hashtab[hash & vcache_hashmask],
1685 vip, vi_hash);
1686 mutex_exit(&vcache_lock);
1687 }
1688 vfs_insmntque(vp, mp);
1689 if ((mp->mnt_iflag & IMNT_MPSAFE) != 0)
1690 vp->v_vflag |= VV_MPSAFE;
1691 vfs_ref(mp);
1692 vfs_unbusy(mp);
1693
1694 /* Finished loading, finalize node. */
1695 mutex_enter(&vcache_lock);
1696 mutex_enter(vp->v_interlock);
1697 VSTATE_CHANGE(vp, VS_LOADING, VS_LOADED);
1698 mutex_exit(&vcache_lock);
1699 mutex_exit(vp->v_interlock);
1700 *vpp = vp;
1701 return 0;
1702 }
1703
1704 /*
1705 * Prepare key change: update old cache nodes key and lock new cache node.
1706 * Return an error if the new node already exists.
1707 */
1708 int
1709 vcache_rekey_enter(struct mount *mp, struct vnode *vp,
1710 const void *old_key, size_t old_key_len,
1711 const void *new_key, size_t new_key_len)
1712 {
1713 uint32_t old_hash, new_hash;
1714 struct vcache_key old_vcache_key, new_vcache_key;
1715 vnode_impl_t *vip, *new_vip;
1716
1717 old_vcache_key.vk_mount = mp;
1718 old_vcache_key.vk_key = old_key;
1719 old_vcache_key.vk_key_len = old_key_len;
1720 old_hash = vcache_hash(&old_vcache_key);
1721
1722 new_vcache_key.vk_mount = mp;
1723 new_vcache_key.vk_key = new_key;
1724 new_vcache_key.vk_key_len = new_key_len;
1725 new_hash = vcache_hash(&new_vcache_key);
1726
1727 new_vip = vcache_alloc();
1728 new_vip->vi_key = new_vcache_key;
1729
1730 /* Insert locked new node used as placeholder. */
1731 mutex_enter(&vcache_lock);
1732 vip = vcache_hash_lookup(&new_vcache_key, new_hash);
1733 if (vip != NULL) {
1734 vcache_dealloc(new_vip);
1735 return EEXIST;
1736 }
1737 SLIST_INSERT_HEAD(&vcache_hashtab[new_hash & vcache_hashmask],
1738 new_vip, vi_hash);
1739
1740 /* Replace old nodes key with the temporary copy. */
1741 vip = vcache_hash_lookup(&old_vcache_key, old_hash);
1742 KASSERT(vip != NULL);
1743 KASSERT(VIMPL_TO_VNODE(vip) == vp);
1744 KASSERT(vip->vi_key.vk_key != old_vcache_key.vk_key);
1745 vip->vi_key = old_vcache_key;
1746 mutex_exit(&vcache_lock);
1747 return 0;
1748 }
1749
1750 /*
1751 * Key change complete: update old node and remove placeholder.
1752 */
1753 void
1754 vcache_rekey_exit(struct mount *mp, struct vnode *vp,
1755 const void *old_key, size_t old_key_len,
1756 const void *new_key, size_t new_key_len)
1757 {
1758 uint32_t old_hash, new_hash;
1759 struct vcache_key old_vcache_key, new_vcache_key;
1760 vnode_impl_t *vip, *new_vip;
1761 struct vnode *new_vp;
1762
1763 old_vcache_key.vk_mount = mp;
1764 old_vcache_key.vk_key = old_key;
1765 old_vcache_key.vk_key_len = old_key_len;
1766 old_hash = vcache_hash(&old_vcache_key);
1767
1768 new_vcache_key.vk_mount = mp;
1769 new_vcache_key.vk_key = new_key;
1770 new_vcache_key.vk_key_len = new_key_len;
1771 new_hash = vcache_hash(&new_vcache_key);
1772
1773 mutex_enter(&vcache_lock);
1774
1775 /* Lookup old and new node. */
1776 vip = vcache_hash_lookup(&old_vcache_key, old_hash);
1777 KASSERT(vip != NULL);
1778 KASSERT(VIMPL_TO_VNODE(vip) == vp);
1779
1780 new_vip = vcache_hash_lookup(&new_vcache_key, new_hash);
1781 KASSERT(new_vip != NULL);
1782 KASSERT(new_vip->vi_key.vk_key_len == new_key_len);
1783 new_vp = VIMPL_TO_VNODE(new_vip);
1784 mutex_enter(new_vp->v_interlock);
1785 VSTATE_ASSERT(VIMPL_TO_VNODE(new_vip), VS_LOADING);
1786 mutex_exit(new_vp->v_interlock);
1787
1788 /* Rekey old node and put it onto its new hashlist. */
1789 vip->vi_key = new_vcache_key;
1790 if (old_hash != new_hash) {
1791 SLIST_REMOVE(&vcache_hashtab[old_hash & vcache_hashmask],
1792 vip, vnode_impl, vi_hash);
1793 SLIST_INSERT_HEAD(&vcache_hashtab[new_hash & vcache_hashmask],
1794 vip, vi_hash);
1795 }
1796
1797 /* Remove new node used as placeholder. */
1798 SLIST_REMOVE(&vcache_hashtab[new_hash & vcache_hashmask],
1799 new_vip, vnode_impl, vi_hash);
1800 vcache_dealloc(new_vip);
1801 }
1802
1803 /*
1804 * Disassociate the underlying file system from a vnode.
1805 *
1806 * Must be called with vnode locked and will return unlocked.
1807 * Must be called with the interlock held, and will return with it held.
1808 */
1809 static void
1810 vcache_reclaim(vnode_t *vp)
1811 {
1812 lwp_t *l = curlwp;
1813 vnode_impl_t *vip = VNODE_TO_VIMPL(vp);
1814 struct mount *mp = vp->v_mount;
1815 uint32_t hash;
1816 uint8_t temp_buf[64], *temp_key;
1817 size_t temp_key_len;
1818 bool recycle;
1819 int error;
1820
1821 KASSERT(VOP_ISLOCKED(vp) == LK_EXCLUSIVE);
1822 KASSERT(mutex_owned(vp->v_interlock));
1823 KASSERT(vrefcnt(vp) != 0);
1824
1825 temp_key_len = vip->vi_key.vk_key_len;
1826 /*
1827 * Prevent the vnode from being recycled or brought into use
1828 * while we clean it out.
1829 */
1830 VSTATE_CHANGE(vp, VS_BLOCKED, VS_RECLAIMING);
1831
1832 /*
1833 * Send NOTE_REVOKE now, before we call VOP_RECLAIM(),
1834 * because VOP_RECLAIM() could cause vp->v_klist to
1835 * become invalid. Don't check for interest in NOTE_REVOKE
1836 * here; it's always posted because it sets EV_EOF.
1837 *
1838 * Once it's been posted, reset vp->v_klist to point to
1839 * our own local storage, in case we were sharing with
1840 * someone else.
1841 */
1842 KNOTE(&vp->v_klist->vk_klist, NOTE_REVOKE);
1843 vp->v_klist = &vip->vi_klist;
1844 mutex_exit(vp->v_interlock);
1845
1846 rw_enter(vp->v_uobj.vmobjlock, RW_WRITER);
1847 mutex_enter(vp->v_interlock);
1848 if ((vp->v_iflag & VI_EXECMAP) != 0) {
1849 cpu_count(CPU_COUNT_EXECPAGES, -vp->v_uobj.uo_npages);
1850 }
1851 vp->v_iflag &= ~(VI_TEXT|VI_EXECMAP);
1852 vp->v_iflag |= VI_DEADCHECK; /* for genfs_getpages() */
1853 mutex_exit(vp->v_interlock);
1854 rw_exit(vp->v_uobj.vmobjlock);
1855
1856 /*
1857 * With vnode state set to reclaiming, purge name cache immediately
1858 * to prevent new handles on vnode, and wait for existing threads
1859 * trying to get a handle to notice VS_RECLAIMED status and abort.
1860 */
1861 cache_purge(vp);
1862
1863 /* Replace the vnode key with a temporary copy. */
1864 if (vip->vi_key.vk_key_len > sizeof(temp_buf)) {
1865 temp_key = kmem_alloc(temp_key_len, KM_SLEEP);
1866 } else {
1867 temp_key = temp_buf;
1868 }
1869 if (vip->vi_key.vk_key_len > 0) {
1870 mutex_enter(&vcache_lock);
1871 memcpy(temp_key, vip->vi_key.vk_key, temp_key_len);
1872 vip->vi_key.vk_key = temp_key;
1873 mutex_exit(&vcache_lock);
1874 }
1875
1876 fstrans_start(mp);
1877
1878 /*
1879 * Clean out any cached data associated with the vnode.
1880 */
1881 error = vinvalbuf(vp, V_SAVE, NOCRED, l, 0, 0);
1882 if (error != 0) {
1883 if (wapbl_vphaswapbl(vp))
1884 WAPBL_DISCARD(wapbl_vptomp(vp));
1885 error = vinvalbuf(vp, 0, NOCRED, l, 0, 0);
1886 }
1887 KASSERTMSG((error == 0), "vinvalbuf failed: %d", error);
1888 KASSERT((vp->v_iflag & VI_ONWORKLST) == 0);
1889 if (vp->v_type == VBLK || vp->v_type == VCHR) {
1890 spec_node_revoke(vp);
1891 }
1892
1893 /*
1894 * Disassociate the underlying file system from the vnode.
1895 * VOP_INACTIVE leaves the vnode locked; VOP_RECLAIM unlocks
1896 * the vnode, and may destroy the vnode so that VOP_UNLOCK
1897 * would no longer function.
1898 */
1899 VOP_INACTIVE(vp, &recycle);
1900 KASSERT(VOP_ISLOCKED(vp) == LK_EXCLUSIVE);
1901 if (VOP_RECLAIM(vp)) {
1902 vnpanic(vp, "%s: cannot reclaim", __func__);
1903 }
1904
1905 KASSERT(vp->v_data == NULL);
1906 KASSERT((vp->v_iflag & VI_PAGES) == 0);
1907
1908 if (vp->v_type == VREG && vp->v_ractx != NULL) {
1909 uvm_ra_freectx(vp->v_ractx);
1910 vp->v_ractx = NULL;
1911 }
1912
1913 if (vip->vi_key.vk_key_len > 0) {
1914 /* Remove from vnode cache. */
1915 hash = vcache_hash(&vip->vi_key);
1916 mutex_enter(&vcache_lock);
1917 KASSERT(vip == vcache_hash_lookup(&vip->vi_key, hash));
1918 SLIST_REMOVE(&vcache_hashtab[hash & vcache_hashmask],
1919 vip, vnode_impl, vi_hash);
1920 mutex_exit(&vcache_lock);
1921 }
1922 if (temp_key != temp_buf)
1923 kmem_free(temp_key, temp_key_len);
1924
1925 /* Done with purge, notify sleepers of the grim news. */
1926 mutex_enter(vp->v_interlock);
1927 vp->v_op = dead_vnodeop_p;
1928 VSTATE_CHANGE(vp, VS_RECLAIMING, VS_RECLAIMED);
1929 vp->v_tag = VT_NON;
1930 mutex_exit(vp->v_interlock);
1931
1932 /*
1933 * Move to dead mount. Must be after changing the operations
1934 * vector as vnode operations enter the mount before using the
1935 * operations vector. See sys/kern/vnode_if.c.
1936 */
1937 vp->v_vflag &= ~VV_ROOT;
1938 vfs_ref(dead_rootmount);
1939 vfs_insmntque(vp, dead_rootmount);
1940
1941 #ifdef PAX_SEGVGUARD
1942 pax_segvguard_cleanup(vp);
1943 #endif /* PAX_SEGVGUARD */
1944
1945 mutex_enter(vp->v_interlock);
1946 fstrans_done(mp);
1947 KASSERT((vp->v_iflag & VI_ONWORKLST) == 0);
1948 }
1949
1950 /*
1951 * Disassociate the underlying file system from an open device vnode
1952 * and make it anonymous.
1953 *
1954 * Vnode unlocked on entry, drops a reference to the vnode.
1955 */
1956 void
1957 vcache_make_anon(vnode_t *vp)
1958 {
1959 vnode_impl_t *vip = VNODE_TO_VIMPL(vp);
1960 uint32_t hash;
1961 bool recycle;
1962
1963 KASSERT(vp->v_type == VBLK || vp->v_type == VCHR);
1964 KASSERT(vp->v_mount == dead_rootmount || fstrans_is_owner(vp->v_mount));
1965 VSTATE_ASSERT_UNLOCKED(vp, VS_ACTIVE);
1966
1967 /* Remove from vnode cache. */
1968 hash = vcache_hash(&vip->vi_key);
1969 mutex_enter(&vcache_lock);
1970 KASSERT(vip == vcache_hash_lookup(&vip->vi_key, hash));
1971 SLIST_REMOVE(&vcache_hashtab[hash & vcache_hashmask],
1972 vip, vnode_impl, vi_hash);
1973 vip->vi_key.vk_mount = dead_rootmount;
1974 vip->vi_key.vk_key_len = 0;
1975 vip->vi_key.vk_key = NULL;
1976 mutex_exit(&vcache_lock);
1977
1978 /*
1979 * Disassociate the underlying file system from the vnode.
1980 * VOP_INACTIVE leaves the vnode locked; VOP_RECLAIM unlocks
1981 * the vnode, and may destroy the vnode so that VOP_UNLOCK
1982 * would no longer function.
1983 */
1984 if (vn_lock(vp, LK_EXCLUSIVE)) {
1985 vnpanic(vp, "%s: cannot lock", __func__);
1986 }
1987 VOP_INACTIVE(vp, &recycle);
1988 KASSERT(VOP_ISLOCKED(vp) == LK_EXCLUSIVE);
1989 if (VOP_RECLAIM(vp)) {
1990 vnpanic(vp, "%s: cannot reclaim", __func__);
1991 }
1992
1993 /* Purge name cache. */
1994 cache_purge(vp);
1995
1996 /* Done with purge, change operations vector. */
1997 mutex_enter(vp->v_interlock);
1998 vp->v_op = spec_vnodeop_p;
1999 vp->v_vflag |= VV_MPSAFE;
2000 mutex_exit(vp->v_interlock);
2001
2002 /*
2003 * Move to dead mount. Must be after changing the operations
2004 * vector as vnode operations enter the mount before using the
2005 * operations vector. See sys/kern/vnode_if.c.
2006 */
2007 vfs_ref(dead_rootmount);
2008 vfs_insmntque(vp, dead_rootmount);
2009
2010 vrele(vp);
2011 }
2012
2013 /*
2014 * Update outstanding I/O count and do wakeup if requested.
2015 */
2016 void
2017 vwakeup(struct buf *bp)
2018 {
2019 vnode_t *vp;
2020
2021 if ((vp = bp->b_vp) == NULL)
2022 return;
2023
2024 KASSERT(bp->b_objlock == vp->v_interlock);
2025 KASSERT(mutex_owned(bp->b_objlock));
2026
2027 if (--vp->v_numoutput < 0)
2028 vnpanic(vp, "%s: neg numoutput, vp %p", __func__, vp);
2029 if (vp->v_numoutput == 0)
2030 cv_broadcast(&vp->v_cv);
2031 }
2032
2033 /*
2034 * Test a vnode for being or becoming dead. Returns one of:
2035 * EBUSY: vnode is becoming dead, with "flags == VDEAD_NOWAIT" only.
2036 * ENOENT: vnode is dead.
2037 * 0: otherwise.
2038 *
2039 * Whenever this function returns a non-zero value all future
2040 * calls will also return a non-zero value.
2041 */
2042 int
2043 vdead_check(struct vnode *vp, int flags)
2044 {
2045
2046 KASSERT(mutex_owned(vp->v_interlock));
2047
2048 if (! ISSET(flags, VDEAD_NOWAIT))
2049 VSTATE_WAIT_STABLE(vp);
2050
2051 if (VSTATE_GET(vp) == VS_RECLAIMING) {
2052 KASSERT(ISSET(flags, VDEAD_NOWAIT));
2053 return EBUSY;
2054 } else if (VSTATE_GET(vp) == VS_RECLAIMED) {
2055 return ENOENT;
2056 }
2057
2058 return 0;
2059 }
2060
2061 int
2062 vfs_drainvnodes(void)
2063 {
2064 int i, gen;
2065
2066 mutex_enter(&vdrain_lock);
2067 for (i = 0; i < 2; i++) {
2068 gen = vdrain_gen;
2069 while (gen == vdrain_gen) {
2070 cv_broadcast(&vdrain_cv);
2071 cv_wait(&vdrain_gen_cv, &vdrain_lock);
2072 }
2073 }
2074 mutex_exit(&vdrain_lock);
2075
2076 if (numvnodes >= desiredvnodes)
2077 return EBUSY;
2078
2079 if (vcache_hashsize != desiredvnodes)
2080 vcache_reinit();
2081
2082 return 0;
2083 }
2084
2085 void
2086 vnpanic(vnode_t *vp, const char *fmt, ...)
2087 {
2088 va_list ap;
2089
2090 #ifdef DIAGNOSTIC
2091 vprint(NULL, vp);
2092 #endif
2093 va_start(ap, fmt);
2094 vpanic(fmt, ap);
2095 va_end(ap);
2096 }
2097
2098 void
2099 vshareilock(vnode_t *tvp, vnode_t *fvp)
2100 {
2101 kmutex_t *oldlock;
2102
2103 oldlock = tvp->v_interlock;
2104 mutex_obj_hold(fvp->v_interlock);
2105 tvp->v_interlock = fvp->v_interlock;
2106 mutex_obj_free(oldlock);
2107 }
2108
2109 void
2110 vshareklist(vnode_t *tvp, vnode_t *fvp)
2111 {
2112 /*
2113 * If two vnodes share klist state, they must also share
2114 * an interlock.
2115 */
2116 KASSERT(tvp->v_interlock == fvp->v_interlock);
2117
2118 /*
2119 * We make the following assumptions:
2120 *
2121 * ==> Some other synchronization is happening outside of
2122 * our view to make this safe.
2123 *
2124 * ==> That the "to" vnode will have the necessary references
2125 * on the "from" vnode so that the storage for the klist
2126 * won't be yanked out from beneath us (the vnode_impl).
2127 *
2128 * ==> If "from" is also sharing, we then assume that "from"
2129 * has the necessary references, and so on.
2130 */
2131 tvp->v_klist = fvp->v_klist;
2132 }
Cache object: ed700c81f2e7513ebb0b37d679c413ad
|