Re: [RFC PATCH 1/7] jump_label: expose queueing API for batched static key updates

[Peter Zijlstra <peterz@infradead.org>]

First strike for not Cc'ing me on all patches in the series :/

On Thu, Mar 05, 2026 at 06:50:04PM -0700, Jim Cromie wrote:
> Currently, `HAVE_JUMP_LABEL_BATCH` provides an architecture-level
> mechanism to defer instruction synchronization (`text_poke_sync()`)
> when patching a sequence of static keys. However, this deferred
> batching capability is not exposed as a public kernel API. Subsystems
> that need to toggle a large number of static keys (e.g.,
> dynamic_debug) currently suffer from O(N) overhead due to repeated
> machine-wide synchronizations (stop_machine).
> 
> This patch introduces a public queueing API to expose this deferred
> synchronization mechanism to the rest of the kernel. This allows
> multiple static keys to be enabled/disabled by queueing their
> architecture-level updates, before applying a single machine-wide
> synchronization barrier after all instructions are modified.
> 
> The new API consists of:
> - static_key_enable_queued(key)
> - static_key_disable_queued(key)
> - static_key_apply_queued() (the global barrier/flush)
> - static_branch_enable_queued(x) / static_branch_disable_queued(x) macros
> 
> NOTES:
> 
> The '_queued' API suffix was chosen to match the underlying
> 'arch_jump_label_transform_queue' and to avoid confusion with the
> existing rate-limited 'static_key_deferred' API.
> 
> Also unify the names under the 'static_key_*' prefix, renaming
> jump_label_apply_queued to static_key_apply_queued (with a
> compatibility macro) for consistency.
> 
> A pr_debug() is added to show the poked addresses, this exposed the
> semi-random ordering coming from dynamic-debug, despite its ordered
> descriptors.
> 
> So x86/kernel/alternatives gets new code to do an insert-sort, by
> memcpy & memmove after appending.  This sorting yields a dramatic IPI
> reduction; a following patch to dynamic-debug uses the API, and
> includes the numbers.
> 

Second strike for doing *WAAAY* to many things in one patch.

> +EXPORT_SYMBOL_GPL(static_key_enable_queued);
> +EXPORT_SYMBOL_GPL(static_key_disable_queued);
> +EXPORT_SYMBOL_GPL(static_key_apply_queued);

Third strike for exposing this to modules; flipping a ton of keys is
*not* something sensible.

> +	pr_debug("incoming addr=%px, current_qlen=%d\n",
> +		 addr, text_poke_array.nr_entries);
> +

And seriously, what?!