Re: [linux-next:master] [slab] db93cdd664: BUG:kernel_NULL_pointer_dereference,address

[Vlastimil Babka <vbabka@suse.cz>]

On 9/19/25 20:31, Alexei Starovoitov wrote:
> On Fri, Sep 19, 2025 at 8:01 AM Suren Baghdasaryan <surenb@google.com> wrote:
>>
>> >
>> > I would not. I think adding 'boot or not' logic to these two
>> > will muddy the waters and will make the whole slab/page_alloc/memcg
>> > logic and dependencies between them much harder to follow.
>> > I'd either add a comment to alloc_slab_obj_exts() explaining
>> > what may happen or add 'boot or not' check only there.
>> > imo this is a niche, rare and special.
>>
>> Ok, comment it is then.
>> Will you be sending a new version or Vlastimil will be including that
>> in his fixup?
> 
> Whichever way. I can, but so far Vlastimil phrasing of comments
> were much better than mine :) So I think he can fold what he prefers.

I'm adding this. Hopefully we'll be able to make sheaves the only percpu
caching layer in SLUB in the (near) future, and then requirement for
cmpxchg16b for allocations will be gone.

diff --git a/mm/slub.c b/mm/slub.c
index 9f1054f0b9ca..f9f7f3942074 100644
--- a/mm/slub.c
+++ b/mm/slub.c
@@ -2089,6 +2089,13 @@ int alloc_slab_obj_exts(struct slab *slab, struct kmem_cache *s,
        gfp &= ~OBJCGS_CLEAR_MASK;
        /* Prevent recursive extension vector allocation */
        gfp |= __GFP_NO_OBJ_EXT;
+
+       /*
+        * Note that allow_spin may be false during early boot and its
+        * restricted GFP_BOOT_MASK. Due to kmalloc_nolock() only supporting
+        * architectures with cmpxchg16b, early obj_exts will be missing for
+        * very early allocations on those.
+        */
        if (unlikely(!allow_spin)) {
                size_t sz = objects * sizeof(struct slabobj_ext);