2010-05-20 05:43:06

by Tao Guo

Subject: [PATCH] nfs: avoid possible deadlock bug when setting up delegation.

Unlock clp->cl_lock before taking inode->i_lock when setting up
delegation to avoid possible deadlock bug.

Signed-off-by: Tao Guo <[email protected]>
---
fs/nfs/delegation.c | 4 +++-
1 files changed, 3 insertions(+), 1 deletions(-)

diff --git a/fs/nfs/delegation.c b/fs/nfs/delegation.c
index ea61d26..0983930 100644
--- a/fs/nfs/delegation.c
+++ b/fs/nfs/delegation.c
@@ -233,6 +233,7 @@ int nfs_inode_set_delegation(struct inode *inode, struct rpc_cred *cred, struct
if (memcmp(&delegation->stateid, &old_delegation->stateid,
sizeof(old_delegation->stateid)) == 0 &&
delegation->type == old_delegation->type) {
+ spin_unlock(&clp->cl_lock);
goto out;
}
/*
@@ -245,6 +246,7 @@ int nfs_inode_set_delegation(struct inode *inode, struct rpc_cred *cred, struct
if (delegation->type <= old_delegation->type) {
freeme = delegation;
delegation = NULL;
+ spin_unlock(&clp->cl_lock);
goto out;
}
freeme = nfs_detach_delegation_locked(nfsi, NULL, clp);
@@ -253,6 +255,7 @@ int nfs_inode_set_delegation(struct inode *inode, struct rpc_cred *cred, struct
nfsi->delegation_state = delegation->type;
rcu_assign_pointer(nfsi->delegation, delegation);
delegation = NULL;
+ spin_unlock(&clp->cl_lock);

/* Ensure we revalidate the attributes and page cache! */
spin_lock(&inode->i_lock);
@@ -260,7 +263,6 @@ int nfs_inode_set_delegation(struct inode *inode, struct rpc_cred *cred, struct
spin_unlock(&inode->i_lock);

out:
- spin_unlock(&clp->cl_lock);
if (delegation != NULL)
nfs_free_delegation(delegation);
if (freeme != NULL)
--
1.6.3.3

2010-05-20 12:57:24

by Trond Myklebust

Subject: Re: [PATCH] nfs: avoid possible deadlock bug when setting up delegation.

On Thu, 2010-05-20 at 13:42 +0800, Tao Guo wrote:
> Unlock clp->cl_lock before taking inode->i_lock when setting up
> delegation to avoid possible deadlock bug.

What deadlock? inode->i_lock should always be the innermost lock, so it
should always be safe to take it while already holding the nfs_client
lock.

If somebody is taking the nfs_client lock while holding the inode lock,
then that would be the deadlock bug that needs to be fixed.

Trond