Re: [PATCH v2 1/1] alpha: Add support for HAVE_ARCH_JUMP_LABEL

[Matt Turner <mattst88@gmail.com>]

On Wed, Feb 25, 2026 at 12:02 PM Magnus Lindholm <linmag7@gmail.com> wrote:
> Implement static key (jump label) support for Alpha.

Sorry for the slow review. This looks correct to me. I applied it to
your for-next (the arch/alpha/Kconfig hunk needs a trivial rebase) and
it builds cleanly with W=1, CONFIG_JUMP_LABEL=y and
CONFIG_STATIC_KEYS_SELFTEST=y. I have not booted it. A few comments,
none of them blocking:

> +   select HAVE_ARCH_JUMP_LABEL

Please keep the select list sorted.

> +static inline void alpha_patch_text(u32 *site, u32 insn)
> +{
> +   WRITE_ONCE(*site, insn);
> +   flush_icache_range((unsigned long)site, (unsigned long)site + sizeof(*site));

On SMP flush_icache_range() is smp_imb(), so this is one
on_each_cpu() IPI round per patched site. A key with many sites, or
a module load, turns into that many broadcasts. That is fine as a
first version, but it would be worth a follow-up that implements
arch_jump_label_transform_queue()/_apply() (HAVE_JUMP_LABEL_BATCH)
so that a key update does a single imb at the end.

> +   if (disp < -(1L << 20) || disp > ((1L << 20) - 1)) {
> +           WARN_ON_ONCE(1);
> +           disp = 0;
> +   }

With disp = 0 we write a branch to the next instruction, so the site
silently behaves as a NOP. I would rather return without patching (or
BUG) than install something that looks valid. The comment about what
other architectures do can go.

> +struct jump_entry {
> +   jump_label_t code;
> +   jump_label_t target;
> +   jump_label_t key;
> +};

Optional: HAVE_ARCH_JUMP_LABEL_RELATIVE would shrink this from 24 to
16 bytes per entry, and the module loader already handles
R_ALPHA_SREL32 and R_ALPHA_SREL64. Fine to leave for later.

Small things:
 - <linux/mutex.h> is unused in jump_label.c.
 - JUMP_LABEL_NOP_SIZE is unused.
 - ALPHA_INSN_NOP has two trailing comments.
 - There is a stray double blank line before the include guard.
 - The commit message could note that this relies on kernel and
   module text being writable (no STRICT_KERNEL_RWX on alpha).

Did you run CONFIG_STATIC_KEYS_SELFTEST, and load and unload a module
that carries static keys?

Reviewed-by: Matt Turner <mattst88@gmail.com>