Re: 3.0.0-rc2-git1 -- BUG: sleeping function called from invalid context at mm/slub.c:847

[John Johansen <john.johansen@canonical.com>]

On 06/08/2011 02:34 PM, David Rientjes wrote:
> On Wed, 8 Jun 2011, Matt Mackall wrote:
> 
>>> Not sure why this ever actually worked with apparmor if prepare_creds() 
>>> does an unconditional GFP_KERNEL allocation since this codepath hasn't 
>>> changed in at least a year and we're holding a spinlock from setrlimit.  
>>> John?
>>
>> Probably a lack of people enabling (and using!) both apparmor and
>> might_sleep. I don't this would be caught by a randconfig boot test.
>>
> 
> Right, CONFIG_DEBUG_SPINLOCK_SLEEP isn't enabled by default even though 
> CONFIG_DEBUG_KERNEL is.  We should probably just allow prepare_creds() to 
> take a gfp_t argument just like security_prepare_creds() and change 
> existing callers to use GFP_KERNEL with the exception of those using 
> setrlimit where we're always holding the spinlock.  
> 
> Documentation/security/credentials.txt says this:
> 
> 	To alter the current process's credentials, a function should first prepare a
> 	new set of credentials by calling:
> 
>         	struct cred *prepare_creds(void);   
> 
> 	this locks current->cred_replace_mutex and then allocates and constructs a
> 	duplicate of the current process's credentials, returning with the mutex still
> 	held if successful.  It returns NULL if not successful (out of memory).
> 
> although that mutex doesn't exist.  David, any downsides to passing the 
> gfp_t into prepare_creds()?

Well it certainly isn't needed for the apparmor case, as the bug is being
triggered by how apparmor handles policy replacement, and we have a means
of handling that for atomic contexts.