Re: [PATCH] nfsd: register pernet ops last, unregister first

[Chuck Lever <chuck.lever@oracle.com>]

> On Feb 3, 2021, at 12:52 PM, Bruce Fields <bfields@fieldses.org> wrote:
> 
> On Wed, Feb 03, 2021 at 04:45:56PM +0000, Chuck Lever wrote:
>> Hi Bruce-
>> 
>> 
>>> On Feb 3, 2021, at 11:42 AM, J. Bruce Fields <bfields@fieldses.org> wrote:
>>> 
>>> From: "J. Bruce Fields" <bfields@redhat.com>
>>> 
>>> These pernet operations may depend on stuff set up or torn down in the
>>> module init/exit functions.  And they may be called at any time in
>>> between.  So it makes more sense for them to be the last to be
>>> registered in the init function, and the first to be unregistered in the
>>> exit function.
>>> 
>>> In particular, without this, the drc slab is being destroyed before all
>>> the per-net drcs are shut down, resulting in an "Objects remaining in
>>> nfsd_drc on __kmem_cache_shutdown()" warning in exit_nfsd.
>>> 
>>> Reported-by: Zhi Li <yieli@redhat.com>
>>> Fixes: 3ba75830ce17 "nfsd4: drc containerization"
>>> Signed-off-by: J. Bruce Fields <bfields@redhat.com>
>> 
>> I can't tell how urgent this is. Does it belong in 5.11-rc?
> 
> I dunno, I wonder what happens when you try to write to and then free a
> bunch of objects that were allocated from a slab that no longer exists.
> 
> But, it's triggered by unloading nfsd, and I find it hard to be super
> concerned about module unloading bugs (does anyone actually *need* to
> unload the nfsd module?).

This doesn't sound urgent, then. My initial 5.12 changeset is closed,
so I'll send this in the first 5.12-rc, unless someone grumbles.


--
Chuck Lever