Re: [PATCH] bpf: Convert lpm_trie::lock to 'raw_spinlock_t'

[Alexei Starovoitov]

On Sat, Nov 16, 2024 at 8:15 AM Thomas Weißschuh <linux@xxxxxxxxxxxxxx> wrote:
>
> On 2024-11-16 08:01:49-0800, Alexei Starovoitov wrote:
> > On Sat, Nov 16, 2024 at 1:21 AM Sebastian Andrzej Siewior
> > <bigeasy@xxxxxxxxxxxxx> wrote:
> > >
> > > On 2024-11-15 23:29:31 [+0100], Thomas Gleixner wrote:
> > > > IIRC, BPF has it's own allocator which can be used everywhere.
> > >
> > > Thomas Weißschuh made something. It appears to work. Need to take a
> > > closer look.
> >
> > Any more details?
> > bpf_mem_alloc is a stop gap.
>
> It is indeed using bpf_mem_alloc.
> It is a fairly straightforward conversion, using one cache for
> intermediate and one for non-intermediate nodes.

Sounds like you're proposing to allocate two lpm specific bpf_ma-s ?
Just use bpf_global_ma.
More ma-s means more memory pinned in bpf specific freelists.
That's the main reason to teach slub and page_alloc about bpf requirements.
All memory should be shared by all subsystems.
Custom memory pools / freelists, whether it's bpf, networking
or whatever else, is a pain point for somebody.
The kernel needs to be optimal for all use cases.

> I'll try to send it early next week.

Looking forward.

> > As Vlastimil Babka suggested long ago:
> > https://lwn.net/Articles/974138/
> > "...next on the target list is the special allocator used by the BPF
> > subsystem. This allocator is intended to succeed in any calling
> > context, including in non-maskable interrupts (NMIs). BPF maintainer
> > Alexei Starovoitov is evidently in favor of this removal if SLUB is
> > able to handle the same use cases..."
> >
> > Here is the first step:
> > https://lore.kernel.org/bpf/20241116014854.55141-1-alexei.starovoitov@xxxxxxxxx/