On Wed, Jul 08, 2026 at 10:27:54PM +0900, Harry Yoo wrote:
>
>
> On 7/8/26 8:16 PM, Pedro Falcato wrote:
> > On Wed, Jul 08, 2026 at 10:30:50AM +0200, Paolo Abeni wrote:
> >> On 7/2/26 7:07 PM, Pedro Falcato wrote:> @@ -586,6 +586,8 @@ struct
> >> sk_buff *napi_build_skb(void *data, unsigned int frag_size)
> >>> }
> >>> EXPORT_SYMBOL(napi_build_skb);
> >>>
> >>> +static kmem_buckets *skb_data_buckets __ro_after_init;
> >>> +
> >>> static void *kmalloc_pfmemalloc(size_t obj_size, gfp_t flags, int node)
> >>> {
> >>> if (!gfp_pfmemalloc_allowed(flags))
> >>> @@ -593,7 +595,8 @@ static void *kmalloc_pfmemalloc(size_t obj_size,
> >>> gfp_t flags, int node)
> >>> if (!obj_size)
> >>> return kmem_cache_alloc_node(net_hotdata.skb_small_head_cache,
> >>> flags, node);
> >>> - return kmalloc_node_track_caller(obj_size, flags, node);
> >>> + return kmem_buckets_alloc_node_track_caller(skb_data_buckets, obj_size,
> >>> + flags, node);
> >>
> >> Sashiko noted that some drivers may require GFP_DMA buckets, and the
> >> above may break them:
> >>
> >> https://sashiko.dev/#/patchset/20260702170728.168755-1-pfalcato%40suse.de
> >
> > Oh, this is really awkward. Adding linux-mm and slab maintainers for input
> > here.
> >
> > Considering the current slab bucketing does not seem to duplicate DMA or
> > CGROUP caches, could it make sense to duplicate those as well?
>
> Could we specify what kmalloc types the user needs when creating
> kmem_buckets and duplicate caches for the requested kmalloc types only?
Perhaps. But do the users themselves know? alloc_skb() allows users to specify
random __GFP flags. We're bound to see some random caller do
alloc_skb(__GFP_ACCOUNT) ;)
In all honesty, I'm not quite sure what the best way forward here is. The most
transparent way is to bucket those other kmalloc types as well, but that might
very trivially result in a lot more caches (and possibly memory usage) for no
great reason. So perhaps specifying caches might do.
--
Pedro
