Re: [PATCH] better locking in fs/proc/generic.c (bug 1552)

[Andrew Morton <akpm@osdl.org>]

Nathan Lynch <nathanl@austin.ibm.com> wrote:
>
> Hi-
> 
> I have been experiencing various oopses in the procfs code when rapidly 
> adding and removing /proc entries.

I doubt if this particularly surprises anyone.

>  It seems that the main problem is 
> lack of proper locking around code that traverses or modifies the 
> proc_dir_entry tree.  There is also the matter of proc_kill_inodes() 
> clearing the file pointer's f_op, which caused oopses in vfs_read() in 
> my testing.
> 
> I have followed the example of proc_lookup() and used 
> lock_kernel/unlock_kernel around critical sections.  I'm not much of a 
> filesystem person, so I would appreciate a review of the patch from 
> folks more knowledgeable.  I have been testing the code on a couple of 
> dual processor machines (i386 and ppc64) without incident.

The correct lock to use is i_sem on the directory which is subject to
lookup.

> Please see http://bugme.osdl.org/show_bug.cgi?id=1552 for oops traces 
> and a testcase.  Andrew posted a one-liner patch there which was 
> included in mm4 which is also needed.

gargh, I was waiting to hear back on that patch, and there it is.  Durned
bugzilla.  Thanks.