Re: [PATCH 12/15] driver core: Implement tagged directory support for device classes.

Tejun Heo [off-list ref] writes:

quoted

To do that I believe we would need to ensure sysfs does not use 
the inode->i_mutex lock except to keep the VFS layer out.  Allowing us
to safely change the directory structure, without holding it.

I don't think sysfs is depending on i_mutex anymore but I need to go
through the code to make sure.

The vfs still does. So at least for directory tree manipulation  we
need to hold i_mutex before we grab sysfs_mutex.

I think that means we need to unscramble the whole set of locking
order issues.

In lookup we have:
local_vfs_lock -> fs_global_lock

In modifications we have:
fs_global_lock -> local_vfs_lock

Which is the definition of a lock ordering problem.

Currently we play jump through some significant hoops to keep things
in local_vfs_lock -> fs_global_lock order.

If we also take the rename_mutex on directory adds and deletes we
may be able to keep jumping through those hoops.  However I expect
we would be in a much better situation if we could figure out how
to avoid the problem.

It looks like the easy way to handle this is to make the sysfs_dirent
list rcu protected.  Which means we can fix our lock ordering problem
without VFS modifications.  Allowing the locking to always
be: sysfs_mutex ... i_mutex.

After that it would be safe and a good idea to have unshared
inodes between superblocks, just so we don't surprise anyone
making generic VFS assumptions.

Eric