RCU-isms in fs/nfs/delegation.c
Subsystems (2)
I'm trying to redo my NFS RCU warning fixup patch on top of Paul's patches, and I've found a small potential bug: nfs_inode_reclaim_delegation() doesn't use the appropriate accessors/locks to protect NFS_I(inode)->delegation, and nor does it use such to protect *delegation that I can see. It just overwrites the record.
Furthermore, for consistency's sake, it should also protect accesses to delegation->cred within that function.
Trond: can you confirm that both NFS_I(inode)->delegation and delegation->cred should be considered RCU-protected pointers?
David
Quoted 5 lines · David Howells
I'm trying to redo my NFS RCU warning fixup patch on top of Paul's patches, and I've found a small potential bug:
nfs_inode_reclaim_delegation()doesn't use the appropriate accessors/locks to protect NFS_I(inode)->delegation, and nor does it use such to protect *delegation that I can see. It just overwrites the record.
Hmm... Yes, I think that function should probably take the rcu_read_lock(), and then take the delegation->lock before modifying the delegation. Furthermore, it should probably fall back to nfs_inode_set_delegation() in case we race with a delegreturn. See below...
Quoted 5 lines
Furthermore, for consistency's sake, it should also protect accesses to
delegation->credwithin that function.Trond: can you confirm that both NFS_I(inode)->delegation and
delegation->credshould be considered RCU-protected pointers?
With the above changes to nfs_inode_reclaim_delegation() I don't think delegation->cred needs to be RCU-protected.
Cheers
Trond
----------------------------------------------------------------------------------------------------------
NFSv4: Fix the locking in nfs_inode_reclaim_delegation()
From: Trond Myklebust <[email protected]>
Ensure that we correctly rcu-dereference the delegation itself, and that we protect against removal while we're changing the contents.
---
fs/nfs/delegation.c | 37 ++++++++++++++++++++++++------------- 1 files changed, 24 insertions(+), 13 deletions(-)
fs/nfs/delegation.c +24 -13
@@ -129,21 +129,32 @@ again:*/ void nfs_inode_reclaim_delegation (struct inode *inode,struct rpc_cred *cred,struct nfs_openres *res) { -struct nfs_delegation *delegation =NFS_I (inode)->delegation; -struct rpc_cred *oldcred; +struct nfs_delegation *delegation; +struct rpc_cred *oldcred =NULL ; +rcu_read_lock (); + delegation =rcu_dereference (NFS_I (inode)->delegation);if (delegation ==NULL ) -return ; -memcpy (delegation->stateid.data, res->delegation.data, -sizeof (delegation->stateid.data)); - delegation->type = res->delegation_type; - delegation->maxsize = res->maxsize; - oldcred = delegation->cred; - delegation->cred =get_rpccred (cred); -clear_bit (NFS_DELEGATION_NEED_RECLAIM , &delegation->flags); -NFS_I (inode)->delegation_state = delegation->type; -smp_wmb (); -put_rpccred (oldcred); + goto out; +spin_lock (&delegation->lock); +if (delegation->inode !=NULL ) { +memcpy (delegation->stateid.data, res->delegation.data, +sizeof (delegation->stateid.data)); + delegation->type = res->delegation_type; + delegation->maxsize = res->maxsize; + oldcred = delegation->cred; + delegation->cred =get_rpccred (cred); +clear_bit (NFS_DELEGATION_NEED_RECLAIM , &delegation->flags); +NFS_I (inode)->delegation_state = delegation->type; +spin_unlock (&delegation->lock); +put_rpccred (oldcred); + }else { +/* We appear to have raced with a delegation return. */ +spin_unlock (&delegation->lock); +nfs_inode_set_delegation (inode, cred, res); + } +out: +rcu_read_unlock (); }static int nfs_do_return_delegation (struct inode *inode,struct nfs_delegation *delegation,int issync)
fs/nfs/delegation.c again: · line 158
Quoted 5 lines · Trond Myklebust
fs/nfs/delegation.c +4
-0+nfs_inode_set_delegation (inode, cred, res); + } +out: +rcu_read_unlock (); }
I think this is wrong. nfs_inode_set_delegation() may sleep as it calls kmalloc() with GFP_KERNEL - but you still have the RCU read lock held. I think you need to drop the RCU read lock before calling it.
David
Quoted 12 lines · David Howells
fs/nfs/delegation.c again: · line 158
Quoted 5 lines · Trond Myklebust
fs/nfs/delegation.c +4
-0+nfs_inode_set_delegation (inode, cred, res); + } +out: +rcu_read_unlock (); }I think this is wrong.
nfs_inode_set_delegation()may sleep as it callskmalloc()withGFP_KERNEL- but you still have the RCU read lock held. I think you need to drop the RCU read lock before calling it.
Quoted 1 line
David
Agreed. Thanks for the review...
Cheers
Trond
Quoted 1 line · Trond Myklebust
Agreed. Thanks for the review...
The version of the patch I posted just before lunch is hopefully fixed correctly.
David