Re: [PATCH RFC v2 06/20] slab: make percpu sheaves compatible with kmalloc_nolock()/kfree_nolock()

[Vlastimil Babka]

On 1/14/26 14:57, Vlastimil Babka wrote:
> On 1/14/26 00:26, Alexei Starovoitov wrote:
>> On Tue, Jan 13, 2026 at 10:36 AM Sebastian Andrzej Siewior
>> <bigeasy@xxxxxxxxxxxxx> wrote:
>>>
>>> On 2026-01-12 16:17:00 [+0100], Vlastimil Babka wrote:
>>> > --- a/mm/slub.c
>>> > +++ b/mm/slub.c
>>> > @@ -5727,6 +5742,12 @@ void *kmalloc_nolock_noprof(size_t size, gfp_t gfp_flags, int node)
>>> >                */
>>> >               return NULL;
>>> >
>>> > +     ret = alloc_from_pcs(s, alloc_gfp, node);
>>> > +     if (ret)
>>> > +             goto success;
>>>
>>> I'm sorry if I am slow but this actually should actually allow
>>> kmalloc_nolock() allocations on PREEMPT_RT from atomic context. I am
>>> mentioning this because of the patch which removes the nmi+hardirq
>>> condtion (https://lore.kernel.org/all/20260113150639.48407-1-swarajgaikwad1925@xxxxxxxxx)
>> 
>> Right. With sheaves kmalloc_nolock() on RT will be more reliable.
> 
> Yes IIRC Hao Li pointed that out before. We'll be able to remove that
> !preemptible() check that we area about to add by the patch above.
> 
> But I'm not sure we can remove (or "not put back") the "in_nmi() ||
> in_hardirq()" too, because as you said it was added with different reasoning
> initially?

Ah right, it was "copied" from alloc_frozen_pages_nolock_noprof() where it's
explained more, and AFAICS will be still applicable with sheaves. We should
add a comment to kmalloc_nolock() referring to the
alloc_frozen_pages_nolock_noprof() comment...