Messages in this thread | | | Subject | Re: [GIT PULL] Please pull NFS client bugfixes.... | From | Peter Zijlstra <> | Date | Thu, 14 Jan 2010 14:18:47 +0100 |
| |
On Thu, 2010-01-07 at 18:00 -0800, Linus Torvalds wrote: > > On Thu, 7 Jan 2010, Linus Torvalds wrote: > > > > You cannot mmap a directory (and you can't readdir a non-directory), so if > > it's a per-inode NFS lock, then the simplest fix _might_ be to just put > > directory locks in a different lockdep class from non-directory locks. > > That might fix it. > > > > Of course, if it's not a per-inode lock, that doesn't help. > > Hmm. I notice that we already do this in unlock_new_inode() for i_mutex. > And I googled for the problem, and it does indeed seem to be about the > normal i_mutex <-> mmap_sem thing with filldir-vs-mmap. > > Is there perhaps some path where NFS creates new inodes without going > through the normal path? Like the root inode or something? So then you'd > have a root inode with i_mutex annotated as being in the same class as a > regular file, which completes the lockdep chain.
Quite possible, but I got lost trying to find anything like the root inode.
> Or maybe there is something else I'm missing. > > I'm adding Peter and Ingo to the Cc as the "lockdep guys". Maybe they see > what I am missing.
Not really, nfs_fhget() seems to know about directories as it has a S_ISDIR() check of itself before doing unlock_new_inode(), so I would expect unlock_new_inode() to indeed set the file_system_type::i_mutex_dir_key.
| |