svn commit: r356441 - in head/sys: compat/linux geom/journal kern sys
Mateusz Guzik
mjg at FreeBSD.org
Tue Jan 7 15:56:27 UTC 2020
Author: mjg
Date: Tue Jan 7 15:56:24 2020
New Revision: 356441
URL: https://svnweb.freebsd.org/changeset/base/356441
Log:
vfs: reimplement deferred inactive to use a dedicated flag (VI_DEFINACT)
The previous behavior of leaving VI_OWEINACT vnodes on the active list without
a hold count is eliminated. Hold count is kept and inactive processing gets
explicitly deferred by setting the VI_DEFINACT flag. The syncer is then
responsible for vdrop.
Reviewed by: kib (previous version)
Tested by: pho (in a larger patch, previous version)
Differential Revision: https://reviews.freebsd.org/D23036
Modified:
head/sys/compat/linux/linux_stats.c
head/sys/geom/journal/g_journal.c
head/sys/kern/vfs_mount.c
head/sys/kern/vfs_subr.c
head/sys/kern/vfs_syscalls.c
head/sys/sys/mount.h
head/sys/sys/vnode.h
Modified: head/sys/compat/linux/linux_stats.c
==============================================================================
--- head/sys/compat/linux/linux_stats.c Tue Jan 7 15:44:19 2020 (r356440)
+++ head/sys/compat/linux/linux_stats.c Tue Jan 7 15:56:24 2020 (r356441)
@@ -666,7 +666,7 @@ linux_syncfs(struct thread *td, struct linux_syncfs_ar
if ((mp->mnt_flag & MNT_RDONLY) == 0 &&
vn_start_write(NULL, &mp, V_NOWAIT) == 0) {
save = curthread_pflags_set(TDP_SYNCIO);
- vfs_msync(mp, MNT_NOWAIT);
+ vfs_periodic(mp, MNT_NOWAIT);
VFS_SYNC(mp, MNT_NOWAIT);
curthread_pflags_restore(save);
vn_finished_write(mp);
Modified: head/sys/geom/journal/g_journal.c
==============================================================================
--- head/sys/geom/journal/g_journal.c Tue Jan 7 15:44:19 2020 (r356440)
+++ head/sys/geom/journal/g_journal.c Tue Jan 7 15:56:24 2020 (r356441)
@@ -2874,7 +2874,7 @@ g_journal_do_switch(struct g_class *classp)
save = curthread_pflags_set(TDP_SYNCIO);
GJ_TIMER_START(1, &bt);
- vfs_msync(mp, MNT_NOWAIT);
+ vfs_periodic(mp, MNT_NOWAIT);
GJ_TIMER_STOP(1, &bt, "Msync time of %s", mountpoint);
GJ_TIMER_START(1, &bt);
Modified: head/sys/kern/vfs_mount.c
==============================================================================
--- head/sys/kern/vfs_mount.c Tue Jan 7 15:44:19 2020 (r356440)
+++ head/sys/kern/vfs_mount.c Tue Jan 7 15:56:24 2020 (r356441)
@@ -1692,7 +1692,7 @@ dounmount(struct mount *mp, int flags, struct thread *
if (coveredvp != NULL)
vdrop(coveredvp);
- vfs_msync(mp, MNT_WAIT);
+ vfs_periodic(mp, MNT_WAIT);
MNT_ILOCK(mp);
async_flag = mp->mnt_flag & MNT_ASYNC;
mp->mnt_flag &= ~MNT_ASYNC;
Modified: head/sys/kern/vfs_subr.c
==============================================================================
--- head/sys/kern/vfs_subr.c Tue Jan 7 15:44:19 2020 (r356440)
+++ head/sys/kern/vfs_subr.c Tue Jan 7 15:56:24 2020 (r356441)
@@ -217,10 +217,9 @@ static int reassignbufcalls;
SYSCTL_INT(_vfs, OID_AUTO, reassignbufcalls, CTLFLAG_RW | CTLFLAG_STATS,
&reassignbufcalls, 0, "Number of calls to reassignbuf");
-static counter_u64_t free_owe_inact;
-SYSCTL_COUNTER_U64(_vfs, OID_AUTO, free_owe_inact, CTLFLAG_RD, &free_owe_inact,
- "Number of times free vnodes kept on active list due to VFS "
- "owing inactivation");
+static counter_u64_t deferred_inact;
+SYSCTL_COUNTER_U64(_vfs, OID_AUTO, deferred_inact, CTLFLAG_RD, &deferred_inact,
+ "Number of times inactive processing was deferred");
/* To keep more than one thread at a time from running vfs_getnewfsid */
static struct mtx mntid_mtx;
@@ -608,7 +607,7 @@ vntblinit(void *dummy __unused)
vnodes_created = counter_u64_alloc(M_WAITOK);
recycles_count = counter_u64_alloc(M_WAITOK);
recycles_free_count = counter_u64_alloc(M_WAITOK);
- free_owe_inact = counter_u64_alloc(M_WAITOK);
+ deferred_inact = counter_u64_alloc(M_WAITOK);
/*
* Initialize the filesystem syncer.
@@ -3012,6 +3011,39 @@ vrefcnt(struct vnode *vp)
return (vp->v_usecount);
}
+static void
+vdefer_inactive(struct vnode *vp)
+{
+
+ ASSERT_VI_LOCKED(vp, __func__);
+ VNASSERT(vp->v_iflag & VI_OWEINACT, vp,
+ ("%s: vnode without VI_OWEINACT", __func__));
+ VNASSERT(!VN_IS_DOOMED(vp), vp,
+ ("%s: doomed vnode", __func__));
+ if (vp->v_iflag & VI_DEFINACT) {
+ VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
+ vdropl(vp);
+ return;
+ }
+ vp->v_iflag |= VI_DEFINACT;
+ VI_UNLOCK(vp);
+ counter_u64_add(deferred_inact, 1);
+}
+
+static void
+vdefer_inactive_cond(struct vnode *vp)
+{
+
+ VI_LOCK(vp);
+ VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
+ if (VN_IS_DOOMED(vp) ||
+ (vp->v_iflag & VI_OWEINACT) == 0) {
+ vdropl(vp);
+ return;
+ }
+ vdefer_inactive(vp);
+}
+
enum vputx_op { VPUTX_VRELE, VPUTX_VPUT, VPUTX_VUNREF };
/*
@@ -3101,8 +3133,12 @@ vputx(struct vnode *vp, enum vputx_op func)
vinactive(vp);
if (func != VPUTX_VUNREF)
VOP_UNLOCK(vp);
+ vdropl(vp);
+ } else if (vp->v_iflag & VI_OWEINACT) {
+ vdefer_inactive(vp);
+ } else {
+ vdropl(vp);
}
- vdropl(vp);
}
/*
@@ -3257,28 +3293,27 @@ vdrop_deactivate(struct vnode *vp)
("vdrop: vnode already reclaimed."));
VNASSERT((vp->v_iflag & VI_FREE) == 0, vp,
("vnode already free"));
+ VNASSERT((vp->v_iflag & VI_OWEINACT) == 0, vp,
+ ("vnode with VI_OWEINACT set"));
+ VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp,
+ ("vnode with VI_DEFINACT set"));
VNASSERT(vp->v_holdcnt == 0, vp,
("vdrop: freeing when we shouldn't"));
- if ((vp->v_iflag & VI_OWEINACT) == 0) {
- mp = vp->v_mount;
- mtx_lock(&mp->mnt_listmtx);
- if (vp->v_iflag & VI_ACTIVE) {
- vp->v_iflag &= ~VI_ACTIVE;
- TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
- mp->mnt_activevnodelistsize--;
- }
- TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
- mp->mnt_tmpfreevnodelistsize++;
- vp->v_iflag |= VI_FREE;
- vp->v_mflag |= VMP_TMPMNTFREELIST;
- VI_UNLOCK(vp);
- if (mp->mnt_tmpfreevnodelistsize >= mnt_free_list_batch)
- vnlru_return_batch_locked(mp);
- mtx_unlock(&mp->mnt_listmtx);
- } else {
- VI_UNLOCK(vp);
- counter_u64_add(free_owe_inact, 1);
+ mp = vp->v_mount;
+ mtx_lock(&mp->mnt_listmtx);
+ if (vp->v_iflag & VI_ACTIVE) {
+ vp->v_iflag &= ~VI_ACTIVE;
+ TAILQ_REMOVE(&mp->mnt_activevnodelist, vp, v_actfreelist);
+ mp->mnt_activevnodelistsize--;
}
+ TAILQ_INSERT_TAIL(&mp->mnt_tmpfreevnodelist, vp, v_actfreelist);
+ mp->mnt_tmpfreevnodelistsize++;
+ vp->v_iflag |= VI_FREE;
+ vp->v_mflag |= VMP_TMPMNTFREELIST;
+ VI_UNLOCK(vp);
+ if (mp->mnt_tmpfreevnodelistsize >= mnt_free_list_batch)
+ vnlru_return_batch_locked(mp);
+ mtx_unlock(&mp->mnt_listmtx);
}
void
@@ -3629,7 +3664,17 @@ vgonel(struct vnode *vp)
*/
active = vp->v_usecount > 0;
oweinact = (vp->v_iflag & VI_OWEINACT) != 0;
- VI_UNLOCK(vp);
+ /*
+ * If we need to do inactive VI_OWEINACT will be set.
+ */
+ if (vp->v_iflag & VI_DEFINACT) {
+ VNASSERT(vp->v_holdcnt > 1, vp, ("lost hold count"));
+ vp->v_iflag &= ~VI_DEFINACT;
+ vdropl(vp);
+ } else {
+ VNASSERT(vp->v_holdcnt > 0, vp, ("vnode without hold count"));
+ VI_UNLOCK(vp);
+ }
vfs_notify_upper(vp, VFS_NOTIFY_UPPER_RECLAIM);
/*
@@ -3823,8 +3868,10 @@ vn_printf(struct vnode *vp, const char *fmt, ...)
strlcat(buf, "|VI_DOINGINACT", sizeof(buf));
if (vp->v_iflag & VI_OWEINACT)
strlcat(buf, "|VI_OWEINACT", sizeof(buf));
+ if (vp->v_iflag & VI_DEFINACT)
+ strlcat(buf, "|VI_DEFINACT", sizeof(buf));
flags = vp->v_iflag & ~(VI_TEXT_REF | VI_MOUNT | VI_FREE | VI_ACTIVE |
- VI_DOINGINACT | VI_OWEINACT);
+ VI_DOINGINACT | VI_OWEINACT | VI_DEFINACT);
if (flags != 0) {
snprintf(buf2, sizeof(buf2), "|VI(0x%lx)", flags);
strlcat(buf, buf2, sizeof(buf));
@@ -4381,23 +4428,67 @@ vfs_unmountall(void)
unmount_or_warn(rootdevmp);
}
-/*
- * perform msync on all vnodes under a mount point
- * the mount point must be locked.
- */
-void
-vfs_msync(struct mount *mp, int flags)
+static void
+vfs_deferred_inactive(struct vnode *vp, int lkflags)
{
+
+ ASSERT_VI_LOCKED(vp, __func__);
+ VNASSERT((vp->v_iflag & VI_DEFINACT) == 0, vp, ("VI_DEFINACT still set"));
+ if ((vp->v_iflag & VI_OWEINACT) == 0) {
+ vdropl(vp);
+ return;
+ }
+ if (vn_lock(vp, lkflags) == 0) {
+ VI_LOCK(vp);
+ if ((vp->v_iflag & (VI_OWEINACT | VI_DOINGINACT)) == VI_OWEINACT)
+ vinactive(vp);
+ VOP_UNLOCK(vp);
+ vdropl(vp);
+ return;
+ }
+ vdefer_inactive_cond(vp);
+}
+
+static void __noinline
+vfs_periodic_inactive(struct mount *mp, int flags)
+{
struct vnode *vp, *mvp;
+ int lkflags;
+
+ lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
+ if (flags != MNT_WAIT)
+ lkflags |= LK_NOWAIT;
+
+ MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
+ if ((vp->v_iflag & VI_DEFINACT) == 0) {
+ VI_UNLOCK(vp);
+ continue;
+ }
+ vp->v_iflag &= ~VI_DEFINACT;
+ vfs_deferred_inactive(vp, lkflags);
+ }
+}
+
+static inline bool
+vfs_want_msync(struct vnode *vp)
+{
struct vm_object *obj;
+
+ if (vp->v_vflag & VV_NOSYNC)
+ return (false);
+ obj = vp->v_object;
+ return (obj != NULL && vm_object_mightbedirty(obj));
+}
+
+static void __noinline
+vfs_periodic_msync_inactive(struct mount *mp, int flags)
+{
+ struct vnode *vp, *mvp;
+ struct vm_object *obj;
struct thread *td;
int lkflags, objflags;
+ bool seen_defer;
- CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
-
- if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
- return;
-
td = curthread;
lkflags = LK_EXCLUSIVE | LK_INTERLOCK;
@@ -4409,9 +4500,16 @@ vfs_msync(struct mount *mp, int flags)
}
MNT_VNODE_FOREACH_ACTIVE(vp, mp, mvp) {
- obj = vp->v_object;
- if (obj == NULL || !vm_object_mightbedirty(obj)) {
- VI_UNLOCK(vp);
+ seen_defer = false;
+ if (vp->v_iflag & VI_DEFINACT) {
+ vp->v_iflag &= ~VI_DEFINACT;
+ seen_defer = true;
+ }
+ if (!vfs_want_msync(vp)) {
+ if (seen_defer)
+ vfs_deferred_inactive(vp, lkflags);
+ else
+ VI_UNLOCK(vp);
continue;
}
if (vget(vp, lkflags, td) == 0) {
@@ -4422,10 +4520,27 @@ vfs_msync(struct mount *mp, int flags)
VM_OBJECT_WUNLOCK(obj);
}
vput(vp);
+ if (seen_defer)
+ vdrop(vp);
+ } else {
+ if (seen_defer)
+ vdefer_inactive_cond(vp);
}
}
}
+void
+vfs_periodic(struct mount *mp, int flags)
+{
+
+ CTR2(KTR_VFS, "%s: mp %p", __func__, mp);
+
+ if ((mp->mnt_kern_flag & MNTK_NOMSYNC) != 0)
+ vfs_periodic_inactive(mp, flags);
+ else
+ vfs_periodic_msync_inactive(mp, flags);
+}
+
static void
destroy_vpollinfo_free(struct vpollinfo *vi)
{
@@ -4636,7 +4751,7 @@ sync_fsync(struct vop_fsync_args *ap)
* batch. Return them instead of letting them stay there indefinitely.
*/
vnlru_return_batch(mp);
- vfs_msync(mp, MNT_NOWAIT);
+ vfs_periodic(mp, MNT_NOWAIT);
error = VFS_SYNC(mp, MNT_LAZY);
curthread_pflags_restore(save);
vn_finished_write(mp);
Modified: head/sys/kern/vfs_syscalls.c
==============================================================================
--- head/sys/kern/vfs_syscalls.c Tue Jan 7 15:44:19 2020 (r356440)
+++ head/sys/kern/vfs_syscalls.c Tue Jan 7 15:56:24 2020 (r356441)
@@ -129,7 +129,7 @@ kern_sync(struct thread *td)
if ((mp->mnt_flag & MNT_RDONLY) == 0 &&
vn_start_write(NULL, &mp, V_NOWAIT) == 0) {
save = curthread_pflags_set(TDP_SYNCIO);
- vfs_msync(mp, MNT_NOWAIT);
+ vfs_periodic(mp, MNT_NOWAIT);
VFS_SYNC(mp, MNT_NOWAIT);
curthread_pflags_restore(save);
vn_finished_write(mp);
Modified: head/sys/sys/mount.h
==============================================================================
--- head/sys/sys/mount.h Tue Jan 7 15:44:19 2020 (r356440)
+++ head/sys/sys/mount.h Tue Jan 7 15:56:24 2020 (r356441)
@@ -396,7 +396,7 @@ void __mnt_vnode_markerfree_active(struct vno
#define MNTK_UNMOUNTF 0x00000001 /* forced unmount in progress */
#define MNTK_ASYNC 0x00000002 /* filtered async flag */
#define MNTK_SOFTDEP 0x00000004 /* async disabled by softdep */
-#define MNTK_NOMSYNC 0x00000008 /* don't do vfs_msync */
+#define MNTK_NOMSYNC 0x00000008 /* don't do msync */
#define MNTK_DRAINING 0x00000010 /* lock draining is happening */
#define MNTK_REFEXPIRE 0x00000020 /* refcount expiring is happening */
#define MNTK_EXTENDED_SHARED 0x00000040 /* Allow shared locking for more ops */
@@ -903,7 +903,7 @@ int vfs_setopts(struct vfsoptlist *opts, const char *n
const char *value);
int vfs_setpublicfs /* set publicly exported fs */
(struct mount *, struct netexport *, struct export_args *);
-void vfs_msync(struct mount *, int);
+void vfs_periodic(struct mount *, int);
int vfs_busy(struct mount *, int);
int vfs_export /* process mount export info */
(struct mount *, struct export_args *);
Modified: head/sys/sys/vnode.h
==============================================================================
--- head/sys/sys/vnode.h Tue Jan 7 15:44:19 2020 (r356440)
+++ head/sys/sys/vnode.h Tue Jan 7 15:56:24 2020 (r356441)
@@ -242,6 +242,7 @@ struct xvnode {
#define VI_ACTIVE 0x0200 /* This vnode is on the active list */
#define VI_DOINGINACT 0x0800 /* VOP_INACTIVE is in progress */
#define VI_OWEINACT 0x1000 /* Need to call inactive */
+#define VI_DEFINACT 0x2000 /* deferred inactive */
#define VV_ROOT 0x0001 /* root of its filesystem */
#define VV_ISTTY 0x0002 /* vnode represents a tty */
More information about the svn-src-all
mailing list