Re: kobject_init_and_add is easy to misuse

[Greg Kroah-Hartman <gregkh@linuxfoundation.org>]

On Tue, Jun 02, 2020 at 12:54:16PM -0700, James Bottomley wrote:
> On Tue, 2020-06-02 at 19:36 +0200, Greg Kroah-Hartman wrote:
> > On Tue, Jun 02, 2020 at 08:25:14AM -0700, James Bottomley wrote:
> > > On Tue, 2020-06-02 at 05:10 -0700, Matthew Wilcox wrote:
> > > > On Tue, Jun 02, 2020 at 07:50:33PM +0800, Wang Hai wrote:
> > > > > syzkaller reports for memory leak when kobject_init_and_add()
> > > > > returns an error in the function sysfs_slab_add() [1]
> > > > > 
> > > > > When this happened, the function kobject_put() is not called
> > > > > for the corresponding kobject, which potentially leads to
> > > > > memory leak.
> > > > > 
> > > > > This patch fixes the issue by calling kobject_put() even if
> > > > > kobject_init_and_add() fails.
> > > > 
> > > > I think this speaks to a deeper problem with
> > > > kobject_init_and_add()
> > > > -- the need to call kobject_put() if it fails is not readily
> > > > apparent
> > > > to most users.  This same bug appears in the first three users of
> > > > kobject_init_and_add() that I checked --
> > > > arch/ia64/kernel/topology.c
> > > > drivers/firmware/dmi-sysfs.c
> > > > drivers/firmware/efi/esrt.c
> > > > drivers/scsi/iscsi_boot_sysfs.c
> > > > 
> > > > Some do get it right --
> > > > arch/powerpc/kernel/cacheinfo.c
> > > > drivers/gpu/drm/ttm/ttm_bo.c
> > > > drivers/gpu/drm/ttm/ttm_memory.c
> > > > drivers/infiniband/hw/mlx4/sysfs.c
> > > > 
> > > > I'd argue that the current behaviour is wrong,
> > > 
> > > Absolutely agree with this.  We have a big meta pattern here where
> > > we introduce functions with tortuous semantics then someone creates
> > > a checker for the semantics and misuses come crawling out of the
> > > woodwork leading to floods of patches, usually for little or never
> > > used error paths, which really don't buy anything apart from
> > > theoretical correctness.  Just insisting on simple semantics would
> > > have avoided this.
> > 
> > I "introduced" this way back at the end of 2007.  It's not exactly a
> > new function.
> 
> Heh, well, if it never fails, how you handle the failure become
> unimportant semantics ...
> 
> > > >  that kobject_init_and_add() should call kobject_put() if the add
> > > > fails.  This would need a tree-wide audit.  But somebody needs to
> > > > do that anyway because based on my random sampling, half of the
> > > > users currently get it wrong.
> > > 
> > > Well, the semantics of kobject_init() are free on fail, so these
> > > are the ones everyone seems to be using.  The semantics of
> > > kobject_add are put on fail.  The problem is that put on fail isn't
> > > necessarily correct in the kobject_init() case: the release
> > > function may make assumptions about the object hierarchy which
> > > aren't satisfied in the kobject_init() failure case.  This argues
> > > that kobject_init_and_add() can't ever have correct semantics and
> > > we should eliminate it.
> > 
> > At the time, it did reduce common functionality and error handling
> > all into a simpler function.  And, given it's history, it must have
> > somehow worked for the past 12 years or so :)
> 
> Well, like I said, as long as it never fails, no problem.
> 
> It was just Matthew saying "couldn't we make it do kobject_put()
> itself?" that got me thinking that perhaps that wouldn't work with all
> cases.  So now we're discussing failure handling, we're into the
> esoteric rabbit hole case that never happens.
> 
> > Odds are, lots of the callers shouldn't be messing around with
> > kobjects in the first place.  Originally it was only assumed that
> > there would be very few users.  But it has spread to filesystems and
> > firmware subsystems.  Drivers should never use it though, so it's a
> > good hint something is wrong there...
> > 
> > Anyway, patches to fix this up to make a "sane" api for kobjects is
> > always appreciated.  Personally I don't have the time at the moment.
> 
> I think the only way we can make the failure semantics consistent is to
> have the kobject_init() ones (so kfree on failure).  That means for the
> add part, the function would have to unwind everything it did from init
> on so kfree() is still an option.  If people agree, then I can produce
> the patch ... it's just the current drive to transform everyone who's
> doing kfree() into kobject_put() would become wrong ...

Everyone should be putting their kfree into the kobject release anyway,
right?

Anyway, let's see your patch before I start to object further :)

thanks,

greg k-h