Re: [PATCH] mm/migrate: Fix NULL movable_ops if CONFIG_ZSMALLOC=m

[David Hildenbrand]

On 17.08.25 03:37, Huacai Chen wrote:
> On Sun, Aug 17, 2025 at 1:02 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
> > On 16.08.25 18:23, Matthew Wilcox wrote:
> > > On Sat, Aug 16, 2025 at 12:54:52PM +0200, David Hildenbrand wrote:
> > > > +++ b/mm/balloon_compaction.c
> > > > @@ -256,8 +256,10 @@ const struct movable_operations balloon_mops = {
> > > >    static int __init balloon_init(void)
> > > >    {
> > > > -    movable_ops[MOVABLE_BALLOON] = &balloon_mops;
> > > > -    return 0;
> > > > +    int rc;
> > > > +
> > > > +    rc = register_movable_ops(&balloon_mops, PGTY_offline);
> > > > +    return rc;
> > >
> > > Using 'rc' as the name of this variable is an anti-pattern.  All it
> > > tells you is "this is the return value".  Calling it 'err' is far
> > > better because now we know it's an error number (or zero for success,
> > > of course).
> >
> > I know, we all have our things to complain about. Some about Cc: above
> > --, others about the name of error variables :P
> >
> > $ git grep "int rc" | wc -l
> > 12730
> > $ git grep "int ret" | wc -l
> > 80386
> > $ git grep "int error" | wc -l
> > 4349
> > $ git grep "int err " | wc -l
> > 6117
> >
> > > It seems to be a particularly IBM derived antipattern ;-)
> >
> > Careful miser :D
> >
> > > Some internal style guide, perhaps?
> >
> > Kernel-internal style guide maybe ;)
> >
> > > > +void unregister_movable_ops(const struct movable_operations *ops, enum pagetype type)
> > > > +{
> > > > +    switch (type) {
> > > > +    case PGTY_offline:
> > > > +            WARN_ON_ONCE(offline_movable_ops != ops);
> > > > +            offline_movable_ops = NULL;
> > > > +            break;
> > > > +    case PGTY_zsmalloc:
> > > > +            WARN_ON_ONCE(zsmalloc_movable_ops != ops);
> > > > +            zsmalloc_movable_ops = NULL;
> > > > +            break;
> > >
> > > This might be a bit excessive ... just passing the pagetype and not
> > > having the sanity checks should be enough for the tiny number of users
> > > this interface will have.
> >
> > Yeah, no strong opinion, this was a 3 minute hack.
> I have tested your code, everything works well. But if the checking is
> too excessive, can I use a set_movalbe_ops() to replace both
> register_movable_ops() and unregister_movable_ops()? 

Makes sense, given that I don't expect us to have multiple handlers for 
the same type in the near future.

Please do keep a sanity check for replacing valid pointers

if (old_ops && new_ops)
	return -EBUSY;

Thanks!

--
Cheers

David / dhildenb