On 7/13/23 14:03, David Sterba wrote:
> On Thu, Jul 13, 2023 at 04:21:10AM -0700, Christoph Hellwig wrote:
>> On Wed, Jul 12, 2023 at 09:17:12PM +0200, David Sterba wrote:
>> > The slab allocator newly allows to disable merging per-slab (since
>> > commit d0bf7d5759c1 ("mm/slab: introduce kmem_cache flag
>> > SLAB_NO_MERGE")). Set this for all caches in debug build so we can
>> > verify there are no leaks when module gets reloaded.
>>
>> So we're having a discussion on linux-mm wether to just disbale slab
>> merging by default, because it really is a pain. Maybe wait for that
>> to settle before adding per-subsystem hacks for what really is a slab
>> problem?
>
> Yeah I can wait with the patch. That slab merging is considered bad is
> new.
Yeah, I wouldn't say it's universally accepted. But even if we change the
default, it's just a default that distros or users might not follow, so
there's still a space for per-cache enforcement IMHO.
> I remember discussions where Linus and (maybe?) xfs guys argued
> pro/against merging of slabs, where xfs wanted not-merging and had to
> resort to hacks like empty slab constructor that would prevent it. I
> can't find the link but that's base of my reasoning to add a flag
> assuming that merging makes sense by default.
Probably this discussion?
https://lore.kernel.org/all/CA+55aFyepmdpbg9U2Pvp+aHjKmmGCrTK2ywzqfmaOTMXQasYNw@xxxxxxxxxxxxxx/
