[LSF/MM/BPF TOPIC] Reentrant kmalloc usable from any context

[LSF/MM/BPF TOPIC] Reentrant kmalloc usable from any context

[Alexei Starovoitov]

A year ago Vlastimil kicked off a discussion:
"What's next for the SLUB allocator"
https://lwn.net/Articles/974138/

One of the proposed goals was to adopt slub allocator to use cases
where bespoke allocators are currently used either due to
performance requirements or context restrictions.
bpf_mem_alloc, kretprobe's objpool, mempool were mentioned
as initial targets. netpoll should probably be on that list too.

Performance might be addressed by sheaves
while context restrictions are more difficult to solve.
bpf, kretprobe, netpool don't know the context where they
might be called. Currently bpf, kretprobe handler preallocate.
Preallocated pools pin memory to one subsystem make it
unavailable to the rest of the kernel.
This has to be fixed. mm subsystem needs to own, share,
distribute the memory.

Agenda for the discussion:
- discuss and agree on problem statement
  what is being solved, why, requirements of the solution
- discuss pros and cons of the existing design that we
  made a bunch of progress on
- brainstorm next steps

Re: [LSF/MM/BPF TOPIC] Reentrant kmalloc usable from any context

[Alexei Starovoitov]

On Thu, Mar 20, 2025 at 11:32 PM Alexei Starovoitov
<alexei.starovoitov@gmail.com> wrote:
>
> A year ago Vlastimil kicked off a discussion:
> "What's next for the SLUB allocator"
> https://lwn.net/Articles/974138/
>
> One of the proposed goals was to adopt slub allocator to use cases
> where bespoke allocators are currently used either due to
> performance requirements or context restrictions.
> bpf_mem_alloc, kretprobe's objpool, mempool were mentioned
> as initial targets. netpoll should probably be on that list too.
>
> Performance might be addressed by sheaves
> while context restrictions are more difficult to solve.
> bpf, kretprobe, netpool don't know the context where they
> might be called. Currently bpf, kretprobe handler preallocate.
> Preallocated pools pin memory to one subsystem make it
> unavailable to the rest of the kernel.
> This has to be fixed. mm subsystem needs to own, share,
> distribute the memory.
>
> Agenda for the discussion:
> - discuss and agree on problem statement
>   what is being solved, why, requirements of the solution
> - discuss pros and cons of the existing design that we
>   made a bunch of progress on
> - brainstorm next steps

Here are the slides from the talk:
https://github.com/4ast/docs/blob/main/Reentrant%20kmalloc.pdf