Re: [PATCH RFC 0/4] firmware: arm: Add Platform Fault Detection Interface support
From: Jithu Joseph
Date: Fri Nov 20 2026 - 19:44:39 EST
On 11/19/2026 7:36 AM, Ahmed Azeem wrote:
> Hi all,
>
> This series proposes an arm64 driver for the Arm Platform Fault
> Detection
> Interface (PFDI), providing a common Linux interface for privileged
> userspace to discover and invoke firmware diagnostics.
>
> We are posting this series as an RFC to seek community feedback on both
> the proposed Linux interface and the underlying firmware specification
> before the specification is finalized. This series and its validation
> use
> the PFDI 1.0 ALP0 specification [1]. The current public revision is PFDI
> 1.0 BET0 [2]. Feedback on the software integration will help identify
> any
> changes needed to align an upstream implementation with the evolving
> specification.
>
...
>
> I would appreciate feedback on the specification's suitability for Linux
> integration, the proposed per-CPU ioctl ABI, the worker and hotplug
> model,
> and whether this belongs under drivers/firmware. In particular, is this
> the right scope for a common PFDI interface, and should error injection
> be exposed separately from normal diagnostic operations?
>
Thanks for the RFC
I haven't read through the specs or code in great detail, but I thought I would send in my initial thoughts
Most of it comes from my experience developing/up-streaming a similar feature in my prior job at Intel called In Field Scan [1] / [2].
I think creating per-cpu threads just for this feature might be an overkill on a system with large core-counts. I think an alternate approach like below might work for this
1. Drop the per-CPU kthreads and the cpuhp state; run each request on the target CPU when it arrives ... something like
cpus_read_lock();
if (!cpu_online(cpu)
ret = -ENODEV;
else
ret = work_on_cpu(cpu, pfdi_op_fn, &req); /* or smp_call_on_cpu() */
cpus_read_unlock();
(IFS used an alternate cpu stopper mechanism as it had the additional responsibility of synchronizing sibling threads and related timing etc ... it might not be needed here )
2. Use one device with the CPU passed in the request, not <numcpu> device nodes. Use a single /dev/pfdi (or /dev/cpu_pfdi) and add a cpu field to each ioctl argument struct. That removes:
- the per-CPU miscdevice and pfdi_dev_data
- the per-CPU register/deregister on hotplug etc
(I will be away next week ...so my responses will be delayed)
[1] https://github.com/torvalds/linux/blob/master/drivers/platform/x86/intel/ifs/ifs.h
[2] https://www.kernel.org/doc/html/latest/arch/x86/ifs.html
Thanks
Jithu