Re: [PATCH] bpf: local_storage: avoid redundant IRQ save on bucket locks

[Amery Hung]

On Wed, Sep 16, 2026 at 11:22 AM Kumar Kartikeya Dwivedi
<memxor@xxxxxxxxx> wrote:
>
> +Cc Amery
>
> On Wed Sep 16, 2026 at 8:10 PM CEST, Usama Arif wrote:
> > bpf_local_storage_update() takes the map bucket lock while holding
> > local_storage->lock. bpf_selem_unlink_map() does the same; its only
> > caller holds local_storage->lock. The outer lock is acquired with
> > raw_res_spin_lock_irqsave(), so interrupts are already disabled at both
> > sites.
> >
> > Using raw_res_spin_lock_irqsave() for the nested lock saves the already
> > disabled IRQ state and issues another IRQ disable. The matching unlock
> > tests that saved state before leaving interrupts disabled. On x86-64,
> > this adds a pushfq/popq/cli sequence and a test/branch around an
> > unreachable sti to each acquisition.
> >
> > Use raw_res_spin_lock() and raw_res_spin_unlock() instead. They retain
> > preemption nesting, memory ordering and resilient-lock bookkeeping. The
> > outer unlock remains responsible for restoring the caller's IRQ state.
> >
> > In the tested clang x86-64 build, this removes five executed instructions
> > from each uncontended nested acquisition. It also shrinks
> > bpf_local_storage_update() from 1732 to 1702 bytes and bpf_selem_unlink()
> > from 1030 to 992 bytes. The affected paths are updates that add or replace
> > an element in existing owner storage and successful unlinks.
> >
> > Document the owner-lock requirement of bpf_selem_unlink_map() and assert
> > that interrupts are disabled.
> >
> > Signed-off-by: Usama Arif <usama.arif@xxxxxxxxx>
> > ---
>
> Makes sense. But did you observe any measurable improvement with this change?
>

Same question, but the change looks right to me.

> > [...]