Re: kobject ->k_name memory leak

[Greg KH <gregkh@suse.de>]

On Sat, Dec 22, 2007 at 03:07:21PM +0300, Alexey Dobriyan wrote:
> On Thu, Dec 20, 2007 at 02:04:56PM -0800, Greg KH wrote:
> > On Sat, Dec 15, 2007 at 04:19:16PM +0100, Kay Sievers wrote:
> > > On Sat, 2007-12-15 at 16:34 +0300, Alexey Dobriyan wrote:
> > > > On Fri, Dec 14, 2007 at 01:48:23PM -0800, Greg KH wrote:
> > > > > On Mon, Dec 03, 2007 at 01:25:51PM -0800, Greg KH wrote:
> > > > > > On Tue, Dec 04, 2007 at 12:09:59AM +0300, Alexey Dobriyan wrote:
> > > > > > > On Mon, Dec 03, 2007 at 12:47:16PM -0800, Greg KH wrote:
> > > > > > > > On Mon, Dec 03, 2007 at 12:26:07PM +0300, Alexey Dobriyan wrote:
> > > > > > > > > Hi, Greg!
> > > > > > > > > 
> > > > > > > > > Commit ce2c9cb0259acd2aed184499ebe41ab00da13b25 aka
> > > > > > > > > "kobject: remove the static array for the name" introduced memory leak
> > > > > > > > > of a module name after modprobe/rmmod. Apparently for modules ->release
> > > > > > > > > callback is NULL.
> > > > > > > > > 
> > > > > > > > > kobject_cleanup: ->release = 00000000, name = 'foo_sysctl'
> > > > > > > > > Pid: 1927, comm: rmmod Not tainted 2.6.24-rc3-e1cca7e8d484390169777b423a7fe46c7021fec1 #5
> > > > > > > > >  [<c10d4a58>] kobject_cleanup+0xb8/0xc0
> > > > > > > > >  [<c10d4a60>] kobject_release+0x0/0x10
> > > > > > > > >  [<c10d587b>] kref_put+0x2b/0xa0
> > > > > > > > >  [<c11dbe85>] _spin_unlock+0x25/0x40
> > > > > > > > >  [<c1045b78>] free_module+0x78/0xd0
> > > > > > > > >  [<c104773f>] sys_delete_module+0x12f/0x1a0
> > > > > > > > 
> > > > > > > > Hm, _which_ kobject associated with a module, there are 3 of them I
> > > > > > > > think :)
> > > > > > > 
> > > > > > > Ouch!
> > > > > > > 
> > > > > > > > They should all have a release function, and if they do not, we think
> > > > > > > > it's a "static" kobject and it is not safe to free that name.
> > > > > > > > 
> > > > > > > > I've been working on cleaning this up a lot in the -mm tree with over 80
> > > > > > > > patches for the kset/kobject apis and interfaces.
> > > > > > > > 
> > > > > > > > But if we have a dynamic kobject, and we aren't freeing it properly,
> > > > > > > > please let me know which one it is and I'll work to fix it for 2.6.24.
> > > > > > > 
> > > > > > > The one which is passed to kobject_set_name() in mod_sysfs_init()..
> > > > > > 
> > > > > > That one should be set to the module_ktype, which is in kernel/params.c,
> > > > > > so the release function there should... oh crap, there is no release
> > > > > > function.  That's a bug.  After I get out of meetings tonight I'll write
> > > > > > up a patch for that, unless someone beats me to it :)
> > > > > 
> > > > > Ok, this is a mess.  We can't really have a release function for this
> > > > > kobject, as the structure it is embedded it has it's own memory
> > > > > management issues.
> > > > > 
> > > > > To fix this properly is going to take some major kobject/module surgery,
> > > > > it's not a simple fix at all.  I'll tackle it for 2.6.25, as it fits in
> > > > > nicely with the other kobject rework that I've already done in the -mm
> > > > > tree.
> > > > > 
> > > > > So, for now, can we just live with this tiny memory leak on module
> > > > > unload?
> > > > 
> > > > For the record, this leak screws any testing one can do wrt module
> > > > unload races. You can't really leave box overnight running
> > > > modprobe/rmmod in a loop, because OOM killer will finally kick in.
> > > > Hey, this is exactly how it was noticed at all.
> > > > 
> > > > > Or is the above trace something that users will see when unloading
> > > > > modules?
> > > > 
> > > > No, it's added debugging.
> > > 
> > > Can't we add an empty release function? It would free the name with the
> > > current logic, right?
> > 
> > Doh, that should be easy.
> > 
> > Alexey, can you see if the patch below fixes this issue for you?
> 
> Yep, it helps!

Great, thanks for testing, I'll send the fix to Linus in a bit.

greg k-h