Re: [RFC PATCH 1/3] arm64: add CPU prefetch and cache modulation sysfs interface

[Breno Leitao <leitao@debian.org>]

On Mon, Aug 17, 2026 at 10:23:33AM +0800, KobaK wrote:
> From: Koba Ko <kobak@nvidia.com>
> 
> Performance characterization on arm64 systems sometimes needs a way to
> inspect and bracket selected implementation-defined prefetch and cache
> controls without exposing raw register contents. Add an RFC-only arm64
> cpumod interface that presents a small set of named, range-checked
> per-CPU sysfs attributes under each CPU device.
> 
> Establish the Kconfig and Makefile plumbing, CPU profile detection for
> Grace and Vera from the target CPU's MIDR, register-field descriptors,
> per-CPU kobject and attribute helpers, locking, and target-CPU read/write
> callbacks. Common attributes are exposed for all supported profiles,
> while Grace- and Vera-specific fields are exposed only for the matching
> profile. Place the option at the end of the top-level Kernel Features
> menu because these implementation controls are not tied to an Arm
> architecture revision.
> 
> Per-CPU cpumod sysfs state follows the CPU lifecycle. Module
> initialization populates subtrees for every online CPU while holding
> cpus_read_lock(), so initial enumeration and CPU hotplug state
> registration are serialized against concurrent topology changes. Offline
> CPUs are intentionally skipped during initial setup because target-CPU
> MIDR detection requires an IPI-capable online CPU; a later online callback
> creates their subtree.
> 
> Profile detection reads MIDR through a synchronous callback on the CPU
> owning the sysfs subtree and returns nonzero only when that dispatch
> fails, so no IPI error can be hidden. After a successful read, retain
> CPUMOD_PROFILE_UNKNOWN as the unsupported-profile sentinel. Subtree
> creation skips that profile during both initial online enumeration and
> later CPUHP online callbacks without blocking module load or CPU hotplug.
> Object and sysfs setup failures remain fatal and are propagated.
> 
> Keep the dynamically allocated kobject state in a private per-CPU pointer
> instead of the CPU device's generic driver-data slot. The online callback
> is idempotent when that pointer already records a subtree. The offline
> callback clears the private pointer before dropping the kobject reference,
> allowing the release callback to retain ownership of the final free.
> 
> Register the dynamic CPUHP state with
> cpuhp_setup_state_nocalls_cpuslocked(). The nocalls form avoids replaying
> startup callbacks for CPUs already initialized by the explicit online
> enumeration. On init failure, release the CPU read lock, destroy all
> subtrees created so far, and return the error. Module exit removes the
> CPUHP state with the matching nocalls helper and tears down all remaining
> per-CPU state across possible CPUs.
> 
> Reviewed-by: Fenghua Yu <fenghuay@nvidia.com>
> Reviewed-by: Tushar Dave <tdave@nvidia.com>
> Reviewed-by: Kai-Heng Feng <kaihengf@nvidia.com>
> Signed-off-by: Koba Ko <kobak@nvidia.com>

Tested-by: Breno Leitao <leitao@debian.org>


> +static struct cpumod_attr _name##_attr = {				\
> +	.kattr = __ATTR(_name, 0644, cpumod_attr_show, cpumod_attr_store), \
> +	.value_offset = offsetof(struct cpumod_subsys, _field),	\
> +	.reg = _reg,						\
> +	.shift = _shift,					\
> +	.field_mask = _field_mask,				\
> +	.max_value = _max_value,				\
> +	.debug_name = _debug_name,				\
> +	.visible_profiles = _visible_profiles,			\
> +}

0644 makes every one of these world-readable, and cpumod_attr_show()
is not a cheap read: it takes subsys->lock and then does a synchronous
smp_call_function_single() to the target CPU, whose callback executes
an implementation-defined mrs.

So any unprivileged user can pick a CPU and hold it under a steady
stream of IPIs just by reading the file in a loop. On a Grace box
running this series:

  # su -s /bin/sh nobody -c 'cat /sys/devices/system/cpu/cpu0/cpumod/pf_dis'
  0