Re: [RFC PATCH] mm/slab: fix a memory leak on kobject_init_and_add() failure

[Vlastimil Babka <vbabka@suse.cz>]

On 10/23/24 11:28, Vlastimil Babka wrote:
> On 10/23/24 11:23, Hyeonggon Yoo wrote:
>> On Wed, Oct 23, 2024 at 1:08 AM Christoph Lameter (Ampere)
>> <cl@gentwo.org> wrote:
>>>
>>> On Tue, 22 Oct 2024, Vlastimil Babka wrote:
>>>
>>> > On 10/21/24 18:27, Christoph Lameter (Ampere) wrote:
>>> > > On Mon, 21 Oct 2024, Vlastimil Babka wrote:
>>> > >
>>> > >> I think the comment "If this function returns an error, kobject_put() must
>>> > >> be called" means that *if* you want to destroy it due to the failure, you
>>> > >> must use kobject_put() and not e.g. kfree(). But IMHO it doesn't mean you
>>> > >> must destroy it because of the kobject_add() failure.
>>> > >
>>> > > Right. The simplest solution is to see the sysfs stuff as optional. If it
>>> >
>>> > To clarify, I only meant the case of boot caches processed for sysfs later.
>>> > I don't think we need to start ignoring all sysfs errors.
>>>
>>> Well not ignoring. Write something to the syslog. So it wont affect slab
>>> operations. /sys support is not critical to the slab subsystem operations
>>> and is often not used at all.
>>>
>>> If its conks out then it should be fixed but it should not impact current
>>> operations. We have had so many issues with sysfs support in the past that
>>> doing so would be wise to avoid future problems.
>> 
>> Both directions look fine to me. Christoph's approach would probably be better
>> for maintainability I think? Failing to create sysfs files is not a
>> critical problem anyway.
> 
> Yeah I guess, if it makes things simpler... Feel free to do that as v2 then!

However we should also make sure aliases (the symlinks) behave sanely and
consistently. Now __kmem_cache_alias() will bail out if sysfs_slab_alias()
fails, for example. To be consistent it should also just log the failure and
proceed. And if the initial mergeable cache fails to create its "unique
name" that aliases symlink to, then all the symlink attempts probably will
also either fail or point to a nonexisting directory. Probably can't be
helped, just make sure it behaves consistently?