2010-06-30 19:12:46

by john stultz

Subject: [PATCH -rt 1/2] Revert select_parent locking fix

Patch against currrent 2.6.33-rt tree.

This reverts 697684652c217b241a07d9e261c7a20e9b072d43 which resolved a
locking issue with select_parent. Unfortunately the fix, which keeps
the entire subchain locked, breaks lockdep.

Reverting so a proper fix can follow.

Signed-off-by: John Stultz <[email protected]>
CC: Thomas Gleixner <[email protected]>
CC: Nick Piggin <[email protected]>
CC: Peter Zijlstra <[email protected]>
CC: John Kacur <[email protected]>
---
fs/dcache.c | 29 +++++++++++++++++------------
1 files changed, 17 insertions(+), 12 deletions(-)

diff --git a/fs/dcache.c b/fs/dcache.c
index 18a3b76..da90157 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -1052,12 +1052,12 @@ resume:

/*
* Descend a level if the d_subdirs list is non-empty.
- * Note that we keep a hold on the parent lock while
- * we descend, so we don't have to reacquire it on
- * ascend.
*/
if (!list_empty(&dentry->d_subdirs)) {
+ spin_unlock(&this_parent->d_lock);
+ spin_release(&dentry->d_lock.dep_map, 1, _RET_IP_);
this_parent = dentry;
+ spin_acquire(&this_parent->d_lock.dep_map, 0, 1, _RET_IP_);
goto repeat;
}

@@ -1071,20 +1071,25 @@ resume:
struct dentry *child;

tmp = this_parent->d_parent;
- child = this_parent;
- next = child->d_u.d_child.next;
+ rcu_read_lock();
spin_unlock(&this_parent->d_lock);
+ child = this_parent;
this_parent = tmp;
+ spin_lock(&this_parent->d_lock);
+ /* might go back up the wrong parent if we have had a rename
+ * or deletion */
+ if (this_parent != child->d_parent ||
+ // d_unlinked(this_parent) || XXX
+ read_seqretry(&rename_lock, seq)) {
+ spin_unlock(&this_parent->d_lock);
+ rcu_read_unlock();
+ goto rename_retry;
+ }
+ rcu_read_unlock();
+ next = child->d_u.d_child.next;
goto resume;
}
-
out:
- /* Make sure we unlock all the way back up the tree */
- while (this_parent != parent) {
- struct dentry *tmp = this_parent->d_parent;
- spin_unlock(&this_parent->d_lock);
- this_parent = tmp;
- }
spin_unlock(&this_parent->d_lock);
if (read_seqretry(&rename_lock, seq))
goto rename_retry;
--
1.6.0.4

2010-06-30 19:12:32

by john stultz

Subject: [PATCH -rt 2/2] Make sure d_kill sets d_parent to null

Patch against currrent 2.6.33-rt tree.

This patch is an alternative fix to the select_parent panic that was
suggested by Nick Piggin.

It resolves the issue by making sure d_parent is set to null so that
the dentry ascending code in select_parent will restart if the child
or next dentry was dkilled while the locks were released.

Credit and thanks to Nick for finding the solution.

Signed-off-by: John Stultz <[email protected]>
CC: Thomas Gleixner <[email protected]>
CC: Nick Piggin <[email protected]>
CC: Peter Zijlstra <[email protected]>
CC: John Kacur <[email protected]>
---
fs/dcache.c | 8 +++++---
1 files changed, 5 insertions(+), 3 deletions(-)

diff --git a/fs/dcache.c b/fs/dcache.c
index da90157..c9d21ae 100644
--- a/fs/dcache.c
+++ b/fs/dcache.c
@@ -225,13 +225,15 @@ static struct dentry *d_kill(struct dentry *dentry)
{
struct dentry *parent;

- list_del(&dentry->d_u.d_child);
- if (dentry->d_parent && dentry != dentry->d_parent)
- spin_unlock(&dentry->d_parent->d_lock);
if (IS_ROOT(dentry))
parent = NULL;
else
parent = dentry->d_parent;
+
+ dentry->d_parent = NULL;
+ list_del(&dentry->d_u.d_child);
+ if (parent)
+ spin_unlock(&parent->d_lock);
/*drops the locks, at that point nobody can reach this dentry */
dentry_iput(dentry);
d_free(dentry);
--
1.6.0.4