Re: [PATCH 0/4] IMA: making i_readcount a first class inode citizen

[Mimi Zohar <zohar@linux.vnet.ibm.com>]

On Fri, 2010-11-05 at 12:28 -0400, J. Bruce Fields wrote: 
> On Fri, Nov 05, 2010 at 07:08:06AM -0400, Mimi Zohar wrote:

> > Right, like the ima_file_check(), which is after the __dentry_open().
> > Al, is it possible to move the break_lease() in may_open() to later?
> 
> That would still leave a race like:
> 
> 			check count
> 	bump count
> 	break lease
> 			set lease
> 
> But we could extend the i_lock to prevent the lease being bumped between
> the two steps on the right-hand side.

The latest i_readcount patchset, i_readcount is atomic and doesn't
require i_lock, at least for IMA.  Have to think about this more ....

> At that point I think we'd be done?  We're assured the count is still
> zero while the lease is added to the inode, so anyone in the process of
> doing an open has yet to reach the break_lease, which will see the newly
> added lease.
> 
> That leaves the problem that leases really should be broken on anything
> that changes the attributes or the dentries pointing to the inode:
> setattr, link, unlink, rename, at least.

For this reason, IMA is now taking i_mutex, preventing file metadata
from changing.

> One approach: add another counter to the inode named disable_leases, and
> have any of those operations do something like:
> 
> 	disable_lease++
> 	break_lease
> 	... do operation ...
> 	disable_lease--
> 
> ?
> 
> --b.

lol, getting i_readcount was was hard enough. 

Mimi