/*-
* Copyright (c) 1997-2011, 2019, 2020 The NetBSD Foundation, Inc.
* All rights reserved.
*
* This code is derived from software contributed to The NetBSD Foundation
* by Jason R. Thorpe of the Numerical Aerospace Simulation Facility,
* NASA Ames Research Center, by Charles M. Hannum, and by Andrew Doran.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
*
* THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
* ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
* TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
* PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
* BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
* CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
* SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
* INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
* CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
* ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
* POSSIBILITY OF SUCH DAMAGE.
*/
/*
* Copyright (c) 1989, 1993
* The Regents of the University of California. All rights reserved.
* (c) UNIX System Laboratories, Inc.
* All or some portions of this file are derived from material licensed
* to the University of California by American Telephone and Telegraph
* Co. or Unix System Laboratories, Inc. and are reproduced herein with
* the permission of UNIX System Laboratories, Inc.
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions
* are met:
* 1. Redistributions of source code must retain the above copyright
* notice, this list of conditions and the following disclaimer.
* 2. Redistributions in binary form must reproduce the above copyright
* notice, this list of conditions and the following disclaimer in the
* documentation and/or other materials provided with the distribution.
* 3. Neither the name of the University nor the names of its contributors
* may be used to endorse or promote products derived from this software
* without specific prior written permission.
*
* THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
* ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
* SUCH DAMAGE.
*
* @(#)vfs_subr.c 8.13 (Berkeley) 4/18/94
*/
/*
* The vnode cache subsystem.
*
* Life-cycle
*
* Normally, there are two points where new vnodes are created:
* VOP_CREATE(9) and VOP_LOOKUP(9). The life-cycle of a vnode
* starts in one of the following ways:
*
* - Allocation, via vcache_get(9) or vcache_new(9).
* - Reclamation of inactive vnode, via vcache_vget(9).
*
* Recycle from a free list, via getnewvnode(9) -> getcleanvnode(9)
* was another, traditional way. Currently, only the draining thread
* recycles the vnodes. This behaviour might be revisited.
*
* The life-cycle ends when the last reference is dropped, usually
* in VOP_REMOVE(9). In such case, VOP_INACTIVE(9) is called to inform
* the file system that vnode is inactive. Via this call, file system
* indicates whether vnode can be recycled (usually, it checks its own
* references, e.g. count of links, whether the file was removed).
*
* Depending on indication, vnode can be put into a free list (cache),
* or cleaned via vcache_reclaim, which calls VOP_RECLAIM(9) to
* disassociate underlying file system from the vnode, and finally
* destroyed.
*
* Vnode state
*
* Vnode is always in one of six states:
* - MARKER This is a marker vnode to help list traversal. It
* will never change its state.
* - LOADING Vnode is associating underlying file system and not
* yet ready to use.
* - LOADED Vnode has associated underlying file system and is
* ready to use.
* - BLOCKED Vnode is active but cannot get new references.
* - RECLAIMING Vnode is disassociating from the underlying file
* system.
* - RECLAIMED Vnode has disassociated from underlying file system
* and is dead.
*
* Valid state changes are:
* LOADING -> LOADED
* Vnode has been initialised in vcache_get() or
* vcache_new() and is ready to use.
* BLOCKED -> RECLAIMING
* Vnode starts disassociation from underlying file
* system in vcache_reclaim().
* RECLAIMING -> RECLAIMED
* Vnode finished disassociation from underlying file
* system in vcache_reclaim().
* LOADED -> BLOCKED
* Either vcache_rekey*() is changing the vnode key or
* vrelel() is about to call VOP_INACTIVE().
* BLOCKED -> LOADED
* The block condition is over.
* LOADING -> RECLAIMED
* Either vcache_get() or vcache_new() failed to
* associate the underlying file system or vcache_rekey*()
* drops a vnode used as placeholder.
*
* Of these states LOADING, BLOCKED and RECLAIMING are intermediate
* and it is possible to wait for state change.
*
* State is protected with v_interlock with one exception:
* to change from LOADING both v_interlock and vcache_lock must be held
* so it is possible to check "state == LOADING" without holding
* v_interlock. See vcache_get() for details.
*
* Reference counting
*
* Vnode is considered active, if reference count (vnode_t::v_usecount)
* is non-zero. It is maintained using: vref(9) and vrele(9), as well
* as vput(9), routines. Common points holding references are e.g.
* file openings, current working directory, mount points, etc.
*
* v_usecount is adjusted with atomic operations, however to change
* from a non-zero value to zero the interlock must also be held.
*/
/*
* There are three lru lists: one holds vnodes waiting for async release,
* one is for vnodes which have no buffer/page references and one for those
* which do (i.e. v_holdcnt is non-zero). We put the lists into a single,
* private cache line as vnodes migrate between them while under the same
* lock (vdrain_lock).
*/
/* Routines having to do with the management of the vnode table. */
/*
* The high bit of v_usecount is a gate for vcache_tryvget(). It's set
* only when the vnode state is LOADED.
* The next bit of v_usecount is a flag for vrelel(). It's set
* from vcache_vget() and vcache_tryvget() whenever the operation succeeds.
*/
#define VUSECOUNT_MASK 0x3fffffff
#define VUSECOUNT_GATE 0x80000000
#define VUSECOUNT_VGET 0x40000000
/*
* Return the current usecount of a vnode.
*/
inline int
vrefcnt(struct vnode *vp)
{
KASSERTMSG(mutex_owned(vp->v_interlock), "at %s:%d", func, line);
if (! VSTATE_VALID(vip->vi_state))
vnpanic(vp, "state is %s at %s:%d",
vstate_name(vip->vi_state), func, line);
KASSERTMSG(mutex_owned(vp->v_interlock), "at %s:%d", func, line);
if (! VSTATE_VALID(vip->vi_state))
vnpanic(vp, "state is %s at %s:%d",
vstate_name(vip->vi_state), func, line);
while (vip->vi_state != VS_LOADED && vip->vi_state != VS_RECLAIMED)
cv_wait(&vp->v_cv, vp->v_interlock);
if (! VSTATE_VALID(vip->vi_state))
vnpanic(vp, "state is %s at %s:%d",
vstate_name(vip->vi_state), func, line);
}
if (! VSTATE_VALID(from))
vnpanic(vp, "from is %s at %s:%d",
vstate_name(from), func, line);
if (! VSTATE_VALID(to))
vnpanic(vp, "to is %s at %s:%d",
vstate_name(to), func, line);
if (vip->vi_state != from)
vnpanic(vp, "from is %s, expected %s at %s:%d\n",
vstate_name(vip->vi_state), vstate_name(from), func, line);
if ((from == VS_LOADED) != gated)
vnpanic(vp, "state is %s, gate %d does not match at %s:%d\n",
vstate_name(vip->vi_state), gated, func, line);
/* Open/close the gate for vcache_tryvget(). */
if (to == VS_LOADED) {
membar_release();
atomic_or_uint(&vp->v_usecount, VUSECOUNT_GATE);
} else {
atomic_and_uint(&vp->v_usecount, ~VUSECOUNT_GATE);
}
atomic_store_relaxed(&vip->vi_state, to);
if (from == VS_LOADING)
cv_broadcast(&vcache_cv);
if (to == VS_LOADED || to == VS_RECLAIMED)
cv_broadcast(&vp->v_cv);
}
/*
* Return the lru list this node should be on.
*/
static vnodelst_t *
lru_which(vnode_t *vp)
{
KASSERT(mutex_owned(vp->v_interlock));
if (vp->v_holdcnt > 0)
return &lru_list[LRU_HOLD];
else
return &lru_list[LRU_FREE];
}
/*
* Put vnode to end of given list.
* Both the current and the new list may be NULL, used on vnode alloc/free.
* Adjust numvnodes and signal vdrain thread if there is work.
*/
static void
lru_requeue(vnode_t *vp, vnodelst_t *listhd)
{
vnode_impl_t *vip;
int d;
/*
* If the vnode is on the correct list, and was put there recently,
* then leave it be, thus avoiding huge cache and lock contention.
*/
vip = VNODE_TO_VIMPL(vp);
if (listhd == vip->vi_lrulisthd &&
(getticks() - vip->vi_lrulisttm) < hz) {
return;
}
mutex_enter(&vdrain_lock);
d = 0;
if (vip->vi_lrulisthd != NULL)
TAILQ_REMOVE(vip->vi_lrulisthd, vip, vi_lrulist);
else
d++;
vip->vi_lrulisthd = listhd;
vip->vi_lrulisttm = getticks();
if (vip->vi_lrulisthd != NULL)
TAILQ_INSERT_TAIL(vip->vi_lrulisthd, vip, vi_lrulist);
else
d--;
if (d != 0) {
/*
* Looks strange? This is not a bug. Don't store
* numvnodes unless there is a change - avoid false
* sharing on MP.
*/
numvnodes += d;
}
if (listhd == &lru_list[LRU_VRELE])
threadpool_schedule_job(threadpool, &vrele_job);
if (d > 0 && numvnodes > desiredvnodes)
threadpool_schedule_job(threadpool, &vdrain_job);
if (d > 0 && numvnodes > desiredvnodes + desiredvnodes / 16)
kpause("vnfull", false, MAX(1, mstohz(10)), &vdrain_lock);
mutex_exit(&vdrain_lock);
}
/*
* Release deferred vrele vnodes for this mount.
* Called with file system suspended.
*/
void
vrele_flush(struct mount *mp)
{
lru_iter_t iter;
vnode_impl_t *vip;
/*
* One pass through the LRU lists to keep the number of allocated
* vnodes below target. Returns true if target met.
*/
static bool
vdrain_one(u_int target)
{
int ix, lists[] = { LRU_FREE, LRU_HOLD };
lru_iter_t iter;
vnode_impl_t *vip;
vnode_t *vp;
struct mount *mp;
KASSERT(mutex_owned(&vdrain_lock));
for (ix = 0; ix < __arraycount(lists); ix++) {
for (vip = lru_iter_first(lists[ix], &iter); vip != NULL;
vip = lru_iter_next(&iter)) {
if (numvnodes < target) {
lru_iter_release(&iter);
return true;
}
vp = VIMPL_TO_VNODE(vip);
/* Probe usecount (unlocked). */
if (vrefcnt(vp) > 0)
continue;
/* Try v_interlock -- we lock the wrong direction! */
if (!mutex_tryenter(vp->v_interlock))
continue;
/* Probe usecount and state. */
if (vrefcnt(vp) > 0 || VSTATE_GET(vp) != VS_LOADED) {
mutex_exit(vp->v_interlock);
continue;
}
mutex_exit(&vdrain_lock);
/*
* Try to drop reference on a vnode. Abort if we are releasing the
* last reference. Note: this _must_ succeed if not the last reference.
*/
static bool
vtryrele(vnode_t *vp)
{
u_int use, next;
membar_release();
for (use = atomic_load_relaxed(&vp->v_usecount);; use = next) {
if (__predict_false((use & VUSECOUNT_MASK) == 1)) {
return false;
}
KASSERT((use & VUSECOUNT_MASK) > 1);
next = atomic_cas_uint(&vp->v_usecount, use, use - 1);
if (__predict_true(next == use)) {
return true;
}
}
}
/*
* vput: unlock and release the reference.
*/
void
vput(vnode_t *vp)
{
int lktype;
/*
* Do an unlocked check of the usecount. If it looks like we're not
* about to drop the last reference, then unlock the vnode and try
* to drop the reference. If it ends up being the last reference
* after all, vrelel() can fix it all up. Most of the time this
* will all go to plan.
*/
if (vrefcnt(vp) > 1) {
VOP_UNLOCK(vp);
if (vtryrele(vp)) {
return;
}
lktype = LK_NONE;
} else {
lktype = VOP_ISLOCKED(vp);
KASSERT(lktype != LK_NONE);
}
mutex_enter(vp->v_interlock);
vrelel(vp, 0, lktype);
}
/*
* Release a vnode from the deferred list.
*/
static void
vrele_deferred(vnode_impl_t *vip)
{
vnode_t *vp;
/*
* First remove the vnode from the vrele list.
* Put it on the last lru list, the last vrele()
* will put it back onto the right list before
* its usecount reaches zero.
*/
TAILQ_REMOVE(vip->vi_lrulisthd, vip, vi_lrulist);
vip->vi_lrulisthd = &lru_list[LRU_HOLD];
vip->vi_lrulisttm = getticks();
TAILQ_INSERT_TAIL(vip->vi_lrulisthd, vip, vi_lrulist);
/*
* Vnode release. If reference count drops to zero, call inactive
* routine and either return to freelist or free to the pool.
*/
static void
vrelel(vnode_t *vp, int flags, int lktype)
{
const bool async = ((flags & VRELEL_ASYNC) != 0);
bool recycle, defer, objlock_held;
u_int use, next;
int error;
objlock_held = false;
retry:
KASSERT(mutex_owned(vp->v_interlock));
if (__predict_false(vp->v_op == dead_vnodeop_p &&
VSTATE_GET(vp) != VS_RECLAIMED)) {
vnpanic(vp, "dead but not clean");
}
/*
* If not the last reference, just unlock and drop the reference count.
*
* Otherwise make sure we pass a point in time where we hold the
* last reference with VGET flag unset.
*/
for (use = atomic_load_relaxed(&vp->v_usecount);; use = next) {
if (__predict_false((use & VUSECOUNT_MASK) > 1)) {
if (objlock_held) {
objlock_held = false;
rw_exit(vp->v_uobj.vmobjlock);
}
if (lktype != LK_NONE) {
mutex_exit(vp->v_interlock);
lktype = LK_NONE;
VOP_UNLOCK(vp);
mutex_enter(vp->v_interlock);
}
if (vtryrele(vp)) {
mutex_exit(vp->v_interlock);
return;
}
next = atomic_load_relaxed(&vp->v_usecount);
continue;
}
KASSERT((use & VUSECOUNT_MASK) == 1);
next = use & ~VUSECOUNT_VGET;
if (next != use) {
next = atomic_cas_uint(&vp->v_usecount, use, next);
}
if (__predict_true(next == use)) {
break;
}
}
membar_acquire();
if (vrefcnt(vp) <= 0 || vp->v_writecount != 0) {
vnpanic(vp, "%s: bad ref count", __func__);
}
/*
* If already clean there is no need to lock, defer or
* deactivate this node.
*/
if (VSTATE_GET(vp) == VS_RECLAIMED) {
if (objlock_held) {
objlock_held = false;
rw_exit(vp->v_uobj.vmobjlock);
}
if (lktype != LK_NONE) {
mutex_exit(vp->v_interlock);
lktype = LK_NONE;
VOP_UNLOCK(vp);
mutex_enter(vp->v_interlock);
}
goto out;
}
/*
* First try to get the vnode locked for VOP_INACTIVE().
* Defer vnode release to vrele task if caller requests
* it explicitly, is the pagedaemon or the lock failed.
*/
defer = false;
if ((curlwp == uvm.pagedaemon_lwp) || async) {
defer = true;
} else if (lktype == LK_SHARED) {
/* Excellent chance of getting, if the last ref. */
error = vn_lock(vp, LK_UPGRADE | LK_RETRY | LK_NOWAIT);
if (error != 0) {
defer = true;
} else {
lktype = LK_EXCLUSIVE;
}
} else if (lktype == LK_NONE) {
/* Excellent chance of getting, if the last ref. */
error = vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_NOWAIT);
if (error != 0) {
defer = true;
} else {
lktype = LK_EXCLUSIVE;
}
}
KASSERT(mutex_owned(vp->v_interlock));
if (defer) {
/*
* Defer reclaim to the vrele task; it's not safe to
* clean it here. We donate it our last reference.
*/
if (lktype != LK_NONE) {
mutex_exit(vp->v_interlock);
VOP_UNLOCK(vp);
mutex_enter(vp->v_interlock);
}
lru_requeue(vp, &lru_list[LRU_VRELE]);
mutex_exit(vp->v_interlock);
return;
}
KASSERT(lktype == LK_EXCLUSIVE);
/* If the node gained another reference, retry. */
use = atomic_load_relaxed(&vp->v_usecount);
if ((use & VUSECOUNT_VGET) != 0) {
goto retry;
}
KASSERT((use & VUSECOUNT_MASK) == 1);
if ((vp->v_iflag & (VI_TEXT|VI_EXECMAP|VI_WRMAP)) != 0 ||
(vp->v_vflag & VV_MAPPED) != 0) {
/* Take care of space accounting. */
if (!objlock_held) {
objlock_held = true;
if (!rw_tryenter(vp->v_uobj.vmobjlock, RW_WRITER)) {
mutex_exit(vp->v_interlock);
rw_enter(vp->v_uobj.vmobjlock, RW_WRITER);
mutex_enter(vp->v_interlock);
goto retry;
}
}
if ((vp->v_iflag & VI_EXECMAP) != 0) {
cpu_count(CPU_COUNT_EXECPAGES, -vp->v_uobj.uo_npages);
}
vp->v_iflag &= ~(VI_TEXT|VI_EXECMAP|VI_WRMAP);
vp->v_vflag &= ~VV_MAPPED;
}
if (objlock_held) {
objlock_held = false;
rw_exit(vp->v_uobj.vmobjlock);
}
/*
* Deactivate the vnode, but preserve our reference across
* the call to VOP_INACTIVE().
*
* If VOP_INACTIVE() indicates that the file has been
* deleted, then recycle the vnode.
*
* Note that VOP_INACTIVE() will not drop the vnode lock.
*/
mutex_exit(vp->v_interlock);
recycle = false;
VOP_INACTIVE(vp, &recycle);
if (!recycle) {
lktype = LK_NONE;
VOP_UNLOCK(vp);
}
mutex_enter(vp->v_interlock);
/*
* Block new references then check again to see if a
* new reference was acquired in the meantime. If
* it was, restore the vnode state and try again.
*/
if (recycle) {
VSTATE_CHANGE(vp, VS_LOADED, VS_BLOCKED);
use = atomic_load_relaxed(&vp->v_usecount);
if ((use & VUSECOUNT_VGET) != 0) {
VSTATE_CHANGE(vp, VS_BLOCKED, VS_LOADED);
goto retry;
}
KASSERT((use & VUSECOUNT_MASK) == 1);
}
/*
* Recycle the vnode if the file is now unused (unlinked).
*/
if (recycle) {
VSTATE_ASSERT(vp, VS_BLOCKED);
KASSERT(lktype == LK_EXCLUSIVE);
/* vcache_reclaim drops the lock. */
lktype = LK_NONE;
vcache_reclaim(vp);
}
KASSERT(vrefcnt(vp) > 0);
KASSERT(lktype == LK_NONE);
out:
for (use = atomic_load_relaxed(&vp->v_usecount);; use = next) {
if (__predict_false((use & VUSECOUNT_VGET) != 0 &&
(use & VUSECOUNT_MASK) == 1)) {
/* Gained and released another reference, retry. */
goto retry;
}
next = atomic_cas_uint(&vp->v_usecount, use, use - 1);
if (__predict_true(next == use)) {
if (__predict_false((use & VUSECOUNT_MASK) != 1)) {
/* Gained another reference. */
mutex_exit(vp->v_interlock);
return;
}
break;
}
}
membar_acquire();
if (VSTATE_GET(vp) == VS_RECLAIMED && vp->v_holdcnt == 0) {
/*
* It's clean so destroy it. It isn't referenced
* anywhere since it has been reclaimed.
*/
vcache_free(VNODE_TO_VIMPL(vp));
} else {
/*
* Otherwise, put it back onto the freelist. It
* can't be destroyed while still associated with
* a file system.
*/
lru_requeue(vp, lru_which(vp));
mutex_exit(vp->v_interlock);
}
}
void
vrele(vnode_t *vp)
{
if (vtryrele(vp)) {
return;
}
mutex_enter(vp->v_interlock);
vrelel(vp, 0, LK_NONE);
}
/*
* Asynchronous vnode release, vnode is released in different context.
*/
void
vrele_async(vnode_t *vp)
{
if (vtryrele(vp)) {
return;
}
mutex_enter(vp->v_interlock);
vrelel(vp, VRELEL_ASYNC, LK_NONE);
}
/*
* Vnode reference, where a reference is already held by some other
* object (for example, a file structure).
*
* NB: lockless code sequences may rely on this not blocking.
*/
void
vref(vnode_t *vp)
{
KASSERT(vrefcnt(vp) > 0);
atomic_inc_uint(&vp->v_usecount);
}
/*
* Page or buffer structure gets a reference.
* Called with v_interlock held.
*/
void
vholdl(vnode_t *vp)
{
/*
* Recycle an unused vnode if caller holds the last reference.
*/
bool
vrecycle(vnode_t *vp)
{
int error __diagused;
mutex_enter(vp->v_interlock);
/* If the vnode is already clean we're done. */
VSTATE_WAIT_STABLE(vp);
if (VSTATE_GET(vp) != VS_LOADED) {
VSTATE_ASSERT(vp, VS_RECLAIMED);
vrelel(vp, 0, LK_NONE);
return true;
}
/* Prevent further references until the vnode is locked. */
VSTATE_CHANGE(vp, VS_LOADED, VS_BLOCKED);
/* Make sure we hold the last reference. */
if (vrefcnt(vp) != 1) {
VSTATE_CHANGE(vp, VS_BLOCKED, VS_LOADED);
mutex_exit(vp->v_interlock);
return false;
}
mutex_exit(vp->v_interlock);
/*
* On a leaf file system this lock will always succeed as we hold
* the last reference and prevent further references.
* On layered file systems waiting for the lock would open a can of
* deadlocks as the lower vnodes may have other active references.
*/
error = vn_lock(vp, LK_EXCLUSIVE | LK_RETRY | LK_NOWAIT);
/*
* Helper for vrevoke() to propagate suspension from lastmp
* to thismp. Both args may be NULL.
* Returns the currently suspended file system or NULL.
*/
static struct mount *
vrevoke_suspend_next(struct mount *lastmp, struct mount *thismp)
{
int error;
if (lastmp == thismp)
return thismp;
if (lastmp != NULL)
vfs_resume(lastmp);
if (thismp == NULL)
return NULL;
do {
error = vfs_suspend(thismp, 0);
} while (error == EINTR || error == ERESTART);
/*
* Eliminate all activity associated with the requested vnode
* and with all vnodes aliased to the requested vnode.
*/
void
vrevoke(vnode_t *vp)
{
struct mount *mp;
vnode_t *vq;
enum vtype type;
dev_t dev;
KASSERT(vrefcnt(vp) > 0);
mp = vrevoke_suspend_next(NULL, vp->v_mount);
mutex_enter(vp->v_interlock);
VSTATE_WAIT_STABLE(vp);
if (VSTATE_GET(vp) == VS_RECLAIMED) {
mutex_exit(vp->v_interlock);
} else if (vp->v_type != VBLK && vp->v_type != VCHR) {
atomic_inc_uint(&vp->v_usecount);
mutex_exit(vp->v_interlock);
vgone(vp);
} else {
dev = vp->v_rdev;
type = vp->v_type;
mutex_exit(vp->v_interlock);
/*
* Eliminate all activity associated with a vnode in preparation for
* reuse. Drops a reference from the vnode.
*/
void
vgone(vnode_t *vp)
{
int lktype;
/*
* Deallocate a vcache node in state VS_LOADING.
*
* vcache_lock held on entry and released on return.
*/
static void
vcache_dealloc(vnode_impl_t *vip)
{
vnode_t *vp;
/*
* Try to get an initial reference on this cached vnode.
* Returns zero on success or EBUSY if the vnode state is not LOADED.
*
* NB: lockless code sequences may rely on this not blocking.
*/
int
vcache_tryvget(vnode_t *vp)
{
u_int use, next;
for (use = atomic_load_relaxed(&vp->v_usecount);; use = next) {
if (__predict_false((use & VUSECOUNT_GATE) == 0)) {
return SET_ERROR(EBUSY);
}
next = atomic_cas_uint(&vp->v_usecount,
use, (use + 1) | VUSECOUNT_VGET);
if (__predict_true(next == use)) {
membar_acquire();
return 0;
}
}
}
/*
* Try to get an initial reference on this cached vnode.
* Returns zero on success and ENOENT if the vnode has been reclaimed.
* Will wait for the vnode state to be stable.
*
* v_interlock locked on entry and unlocked on exit.
*/
int
vcache_vget(vnode_t *vp)
{
int error;
KASSERT(mutex_owned(vp->v_interlock));
/* Increment hold count to prevent vnode from disappearing. */
vp->v_holdcnt++;
VSTATE_WAIT_STABLE(vp);
vp->v_holdcnt--;
/* If this was the last reference to a reclaimed vnode free it now. */
if (__predict_false(VSTATE_GET(vp) == VS_RECLAIMED)) {
if (vp->v_holdcnt == 0 && vrefcnt(vp) == 0)
vcache_free(VNODE_TO_VIMPL(vp));
else
mutex_exit(vp->v_interlock);
return SET_ERROR(ENOENT);
}
VSTATE_ASSERT(vp, VS_LOADED);
error = vcache_tryvget(vp);
KASSERT(error == 0);
mutex_exit(vp->v_interlock);
return 0;
}
/*
* Get a vnode / fs node pair by key and return it referenced through vpp.
*/
int
vcache_get(struct mount *mp, const void *key, size_t key_len,
struct vnode **vpp)
{
int error;
uint32_t hash;
const void *new_key;
struct vnode *vp;
struct vcache_key vcache_key;
vnode_impl_t *vip, *new_vip;
/* If found, take a reference or retry. */
if (__predict_true(vip != NULL)) {
/*
* If the vnode is loading we cannot take the v_interlock
* here as it might change during load (see uvm_obj_setlock()).
* As changing state from VS_LOADING requires both vcache_lock
* and v_interlock it is safe to test with vcache_lock held.
*
* Wait for vnodes changing state from VS_LOADING and retry.
*/
if (__predict_false(vip->vi_state == VS_LOADING)) {
cv_wait(&vcache_cv, &vcache_lock);
mutex_exit(&vcache_lock);
goto again;
}
vp = VIMPL_TO_VNODE(vip);
mutex_enter(vp->v_interlock);
mutex_exit(&vcache_lock);
error = vcache_vget(vp);
if (error == ENOENT)
goto again;
if (error == 0)
*vpp = vp;
KASSERT((error != 0) == (*vpp == NULL));
return error;
}
mutex_exit(&vcache_lock);
/* Allocate and initialize a new vcache / vnode pair. */
error = vfs_busy(mp);
if (error)
return error;
new_vip = vcache_alloc();
new_vip->vi_key = vcache_key;
vp = VIMPL_TO_VNODE(new_vip);
mutex_enter(&vcache_lock);
vip = vcache_hash_lookup(&vcache_key, hash);
if (vip == NULL) {
SLIST_INSERT_HEAD(&vcache_hashtab[hash & vcache_hashmask],
new_vip, vi_hash);
vip = new_vip;
}
/* If another thread beat us inserting this node, retry. */
if (vip != new_vip) {
vcache_dealloc(new_vip);
vfs_unbusy(mp);
goto again;
}
mutex_exit(&vcache_lock);
/* Rekey old node and put it onto its new hashlist. */
vip->vi_key = new_vcache_key;
if (old_hash != new_hash) {
SLIST_REMOVE(&vcache_hashtab[old_hash & vcache_hashmask],
vip, vnode_impl, vi_hash);
SLIST_INSERT_HEAD(&vcache_hashtab[new_hash & vcache_hashmask],
vip, vi_hash);
}
/* Remove new node used as placeholder. */
SLIST_REMOVE(&vcache_hashtab[new_hash & vcache_hashmask],
new_vip, vnode_impl, vi_hash);
vcache_dealloc(new_vip);
}
/*
* Disassociate the underlying file system from a vnode.
*
* Must be called with vnode locked and will return unlocked.
* Must be called with the interlock held, and will return with it held.
*/
static void
vcache_reclaim(vnode_t *vp)
{
lwp_t *l = curlwp;
vnode_impl_t *vip = VNODE_TO_VIMPL(vp);
struct mount *mp = vp->v_mount;
uint32_t hash;
uint8_t temp_buf[64], *temp_key;
size_t temp_key_len;
bool recycle;
int error;
temp_key_len = vip->vi_key.vk_key_len;
/*
* Prevent the vnode from being recycled or brought into use
* while we clean it out.
*/
VSTATE_CHANGE(vp, VS_BLOCKED, VS_RECLAIMING);
/*
* Send NOTE_REVOKE now, before we call VOP_RECLAIM(),
* because VOP_RECLAIM() could cause vp->v_klist to
* become invalid. Don't check for interest in NOTE_REVOKE
* here; it's always posted because it sets EV_EOF.
*
* Once it's been posted, reset vp->v_klist to point to
* our own local storage, in case we were sharing with
* someone else.
*/
KNOTE(&vp->v_klist->vk_klist, NOTE_REVOKE);
vp->v_klist = &vip->vi_klist;
mutex_exit(vp->v_interlock);
/*
* With vnode state set to reclaiming, purge name cache immediately
* to prevent new handles on vnode, and wait for existing threads
* trying to get a handle to notice VS_RECLAIMED status and abort.
*/
cache_purge(vp);
/* Replace the vnode key with a temporary copy. */
if (vip->vi_key.vk_key_len > sizeof(temp_buf)) {
temp_key = kmem_alloc(temp_key_len, KM_SLEEP);
} else {
temp_key = temp_buf;
}
if (vip->vi_key.vk_key_len > 0) {
mutex_enter(&vcache_lock);
memcpy(temp_key, vip->vi_key.vk_key, temp_key_len);
vip->vi_key.vk_key = temp_key;
mutex_exit(&vcache_lock);
}
fstrans_start(mp);
/*
* Clean out any cached data associated with the vnode.
*/
error = vinvalbuf(vp, V_SAVE, NOCRED, l, 0, 0);
if (error != 0) {
if (wapbl_vphaswapbl(vp))
WAPBL_DISCARD(wapbl_vptomp(vp));
error = vinvalbuf(vp, 0, NOCRED, l, 0, 0);
}
KASSERTMSG((error == 0), "vinvalbuf failed: %d", error);
KASSERT((vp->v_iflag & VI_ONWORKLST) == 0);
if (vp->v_type == VBLK || vp->v_type == VCHR) {
spec_node_revoke(vp);
}
/*
* Disassociate the underlying file system from the vnode.
* VOP_INACTIVE leaves the vnode locked; VOP_RECLAIM unlocks
* the vnode, and may destroy the vnode so that VOP_UNLOCK
* would no longer function.
*/
VOP_INACTIVE(vp, &recycle);
KASSERT(VOP_ISLOCKED(vp) == LK_EXCLUSIVE);
if (VOP_RECLAIM(vp)) {
vnpanic(vp, "%s: cannot reclaim", __func__);
}
/* Done with purge, notify sleepers of the grim news. */
mutex_enter(vp->v_interlock);
vp->v_op = dead_vnodeop_p;
VSTATE_CHANGE(vp, VS_RECLAIMING, VS_RECLAIMED);
vp->v_tag = VT_NON;
mutex_exit(vp->v_interlock);
/*
* Move to dead mount. Must be after changing the operations
* vector as vnode operations enter the mount before using the
* operations vector. See sys/kern/vnode_if.c.
*/
vp->v_vflag &= ~VV_ROOT;
vfs_ref(dead_rootmount);
vfs_insmntque(vp, dead_rootmount);
/*
* Disassociate the underlying file system from an open device vnode
* and make it anonymous.
*
* Vnode unlocked on entry, drops a reference to the vnode.
*/
void
vcache_make_anon(vnode_t *vp)
{
vnode_impl_t *vip = VNODE_TO_VIMPL(vp);
uint32_t hash;
bool recycle;
/*
* Disassociate the underlying file system from the vnode.
* VOP_INACTIVE leaves the vnode locked; VOP_RECLAIM unlocks
* the vnode, and may destroy the vnode so that VOP_UNLOCK
* would no longer function.
*/
if (vn_lock(vp, LK_EXCLUSIVE)) {
vnpanic(vp, "%s: cannot lock", __func__);
}
VOP_INACTIVE(vp, &recycle);
KASSERT(VOP_ISLOCKED(vp) == LK_EXCLUSIVE);
if (VOP_RECLAIM(vp)) {
vnpanic(vp, "%s: cannot reclaim", __func__);
}
/*
* Move to dead mount. Must be after changing the operations
* vector as vnode operations enter the mount before using the
* operations vector. See sys/kern/vnode_if.c.
*/
vfs_ref(dead_rootmount);
vfs_insmntque(vp, dead_rootmount);
vrele(vp);
}
/*
* Update outstanding I/O count and do wakeup if requested.
*/
void
vwakeup(struct buf *bp)
{
vnode_t *vp;
if (--vp->v_numoutput < 0)
vnpanic(vp, "%s: neg numoutput, vp %p", __func__, vp);
if (vp->v_numoutput == 0)
cv_broadcast(&vp->v_cv);
}
/*
* Test a vnode for being or becoming dead. Returns one of:
* EBUSY: vnode is becoming dead, with "flags == VDEAD_NOWAIT" only.
* ENOENT: vnode is dead.
* 0: otherwise.
*
* Whenever this function returns a non-zero value all future
* calls will also return a non-zero value.
*/
int
vdead_check(struct vnode *vp, int flags)
{
KASSERT(mutex_owned(vp->v_interlock));
if (! ISSET(flags, VDEAD_NOWAIT))
VSTATE_WAIT_STABLE(vp);
if (VSTATE_GET(vp) == VS_RECLAIMING) {
KASSERT(ISSET(flags, VDEAD_NOWAIT));
return SET_ERROR(EBUSY);
} else if (VSTATE_GET(vp) == VS_RECLAIMED) {
return SET_ERROR(ENOENT);
}
return 0;
}
int
vfs_drainvnodes(void)
{
mutex_enter(&vdrain_lock);
if (!vdrain_one(desiredvnodes)) {
mutex_exit(&vdrain_lock);
return SET_ERROR(EBUSY);
}
mutex_exit(&vdrain_lock);
if (vcache_hashsize != desiredvnodes)
vcache_reinit();
void
vshareklist(vnode_t *tvp, vnode_t *fvp)
{
/*
* If two vnodes share klist state, they must also share
* an interlock.
*/
KASSERT(tvp->v_interlock == fvp->v_interlock);
/*
* We make the following assumptions:
*
* ==> Some other synchronization is happening outside of
* our view to make this safe.
*
* ==> That the "to" vnode will have the necessary references
* on the "from" vnode so that the storage for the klist
* won't be yanked out from beneath us (the vnode_impl).
*
* ==> If "from" is also sharing, we then assume that "from"
* has the necessary references, and so on.
*/
tvp->v_klist = fvp->v_klist;
}
