Re: [PATCH v2 0/3] Refactor __kmem_cache_create() and fix memory leak

[Vlastimil Babka <vbabka@suse.cz>]

On 11/3/22 14:23, Hyeonggon Yoo wrote:
> On Wed, Nov 02, 2022 at 04:53:08PM +0800, Liu Shixin wrote:
>> On 2022/11/2 15:46, Hyeonggon Yoo wrote:
>> > On Mon, Oct 31, 2022 at 09:47:44PM +0800, Liu Shixin wrote:
>> >> I found a memory leak of kobj->name in sysfs_slab_add() which is introduced
>> >> by 80da026a8e5d ("mm/slub: fix slab double-free in case of duplicate sysfs filename").
>> >> Following the rules stated in the comment for kobject_init_and_add():
>> > Thank you for reporting this! Indeed it seems tried to fix double free but
>> > introduced a leak.
>> >
>> >>  If this function returns an error, kobject_put() must be called to
>> >>  properly clean up the memory associated with the object.
>> >>
>> >> We should use kobject_put() to free kobject.
>> > But what to do if a cache is created early and later sysfs_slab_add() failed?
>> > (Which is unlikely on normal condition)
>> >
>> > With this series it introduces use-after-free if sysfs_slab_add() in
>> > slab_sysfs_init() failed.  Should we just call BUG() or something like that?
>>
>> Thanks for your discovery, what I missed.
> 
> You're welcome.
> 
>> I prefer to panic directly, just as create_boot_cache() does.
> 
> IMHO that should be nothing serious. but let's hear maintainers' opinion.
> 
>> Of couse, if you want the system to continue booting, I think it's possible to distinguish them
>> by slab_state.
> 
> I'm afraid to make it more complex :(

+Cc Rasmus who did a recent patch in this area. Thread starts here:
https://lore.kernel.org/all/20221031134747.3049593-1-liushixin2@huawei.com/

As for me, I don't think we should be addung new BUG() or panic() in
general, and especially not for a failing sysfs add. AFAICS
create_boot_cache() might panic, but not because of sysfs, as that's delayed
until slab_sysfs_init() and we don't panic anymore in the latter.

So yeah, it could work to tell sysfs_slab_add() whether it should not do the
kobject_put() as it's a boot cache. The slab_state should work.

> 
>> Looking forward to your advice.
>> Thanks,
>> >
>> >> But we can't simply add kobject_put() since it will free kmem_cache too.
>> >> If we use kobject_put(), we need to skip other release functions.
>> >>
>> >> In this series, We refactor the code to separate sysfs_slab_add() and
>> >> debugfs_slab_add() from __kmem_cache_create(), and then use kobject_put()
>> >> to free kobject in sysfs_slab_add(). This can fix the memory leak of
>> >> kobject->name.
>> >>
>> >> v1->v2: Fix build error reported by kernel test robot <lkp@intel.com>.
>> >>
>> >> Liu Shixin (3):
>> >>   mm/slab_common: Move cache_name to create_cache()
>> >>   mm/slub: Refactor __kmem_cache_create()
>> >>   mm/slub: Fix memory leak of kobj->name in sysfs_slab_add()
>> >>
>> >>  include/linux/slub_def.h | 11 +++++++++
>> >>  mm/slab_common.c         | 44 ++++++++++++++++++----------------
>> >>  mm/slub.c                | 52 ++++++++++------------------------------
>> >>  3 files changed, 48 insertions(+), 59 deletions(-)
>> >>
>> >> -- 
>> >> 2.25.1
>> >>
>> 
>