Re: [PATCH v7 00/11] arm_mpam: Add MPAM-Fb firmware support

[Gavin Shan <gshan@redhat.com>]

On 8/1/26 3:03 AM, Andre Przywara wrote:
> v7 is another quick respin of the MPAM-Fb code, for firmware based MSC
> accesses. This propagates MSC access errors in a function where it was
> missing before, covering more functions that pass their return value
> to userspace (new patch 07/11). Also some bug fixes, formatting
> adjustments and adding the accumulated tags - many thanks to the
> reviewers. Find the changelog below. Based on v7.2-rc1.
> 
> =======================
> The Arm MPAM specification defines Memory System Components (MSCs),
> which are devices that are programmed through an MMIO register frame. In
> some occasions this turned out to be too limiting: the MSC might be
> located behind a separate bus system (for instance inside an on-board
> controller), it might be mapped secure-only, or in a different processor
> socket without direct MMIO mapping. Also the MMIO access might be too slow
> or it would need to be filtered or otherwise access controlled. Finally
> there might be bugs in the MSC integration, which require a mediating
> firmware to be accessible.
> 
> To accommodate all those different use cases, the MPAM-Fb specification
> [1] describes an alternative way to access MSCs. Accesses to an MSC
> would be wrapped in a message and communicated to the system using a
> shared-memory/mailbox system mostly mimicking the Arm SCMI spec.
> For ACPI systems, this would be abstracted through an ACPI PCC channel,
> which provides the shared-memory region and the mailbox trigger. We can
> lean on existing ACPI parsing code to register with these two
> subsystems, but cannot rely on the existing SCMI code in the kernel.
> This means we somewhat need to open code a very simplified SCMI handler,
> which just provides enough functionality for the very basic subset of
> SCMI that the MPAM-Fb spec requires.
> 
> The first seven patches rework all MSC access wrappers to propagate error
> information. Pure MMIO based MSC accesses would never fail, but the
> MPAM-Fb access can go wrong in multiple ways. The patches have been split
> up purely for reviewing reasons, if the number is a problem, we could as
> well squash them. Please note that until the very last patch of this series
> any MSC accesses would always only return 0, it's only the final enablement
> of MPAM-Fb that could possibly introduce errors. Hence all former patches
> can add error handling gradually, those code paths wouldn't be triggered
> before patch 11/11.
> Patch 8/11 solves a nasty problem: At the moment we protect stateful MSC
> register accesses (mon_sel) through a spinlock. Unfortunately the mailbox
> subsystem and the slow nature of the communication through this channel
> forbid MPAM-Fb access in atomic context. So this patch keeps using a
> spinlock for MMIO based accesses, but reverts to a mutex otherwise.
> We just deny taking the lock for MPAM-Fb in atomic context, ideally we
> wouldn't need that (no need to IPI another core when the MSC access does
> not need to be local to one particular core), or we simply deny that part
> of the functionality (access through perf).
> Patch 9/11 adds the code to redirect MSC accesses through the
> PCC shmem/mailbox system.
> Patch 10/11 reworks the error interrupt handler to use a threaded IRQ for
> MPAM-Fb, to be able to do MPAM-Fb MSC accesses inside (which might sleep).
> The final patch 11/11 then adds the code to detect and store the PCC
> channel information from the ACPI tables, and eventually enables
> MPAM-Fb accesses.
> 
> This would enable systems where some MSCs are not accessible via MMIO to
> use those components anyway.
> 
> Please have a look and test!
> 
> Cheers,
> Andre
> 
> [1] https://developer.arm.com/documentation/den0144/latest
> 
> Changes in v7:
> - add tags
> - add new patch to propagate errors in mpam_reprogram_ris_partid()
> - prevent loop when mpam_diable() tries MSC accesses again
> - register MMIO error IRQ handler without IRQF_ONESHOT
> - re-use existing "err" variable instead of declaring "ret"
> - initialise mon_sel_lock later, to wait for interface decision
> - convert timeout units for nominal latency, between us and ms
>   

[...]

> 
> Andre Przywara (11):
>    arm_mpam: let low level MSC accessors return an error
>    arm_mpam: propagate MSC access errors for hw_probe functions
>    arm_mpam: propagate MSC access errors for MBWU counters
>    arm_mpam: propagate MSC access errors for msmon helpers
>    arm_mpam: propagate MSC access errors for __ris_msmon_read()
>    arm_mpam: propagate MSC access errors for state saving function
>    arm_mpam: propagate MSC access errors for mpam_reprogram_ris_partid()
>    arm_mpam: prepare mon_sel locking for MPAM-Fb
>    arm_mpam: add MPAM-Fb MSC firmware access support
>    arm_mpam: change MPAM-Fb error IRQ to use a threaded IRQ handler
>    arm_mpam: detect and enable MPAM-Fb PCC support
> 
>   drivers/resctrl/Makefile        |   2 +-
>   drivers/resctrl/mpam_devices.c  | 820 +++++++++++++++++++++++---------
>   drivers/resctrl/mpam_fb.c       | 250 ++++++++++
>   drivers/resctrl/mpam_internal.h |  62 ++-
>   include/linux/arm_mpam.h        |   2 +-
>   5 files changed, 915 insertions(+), 221 deletions(-)
>   create mode 100644 drivers/resctrl/mpam_fb.c
> 

With this series (manually) applied to v7.2.rc6, the tests for the existing functions
like kunit-tests, L3 cache partitioning, MBW (soft) limiting, llc_occupancy monitor
look fine on NVidia's grace-hopper machine.

Tested-by: Gavin Shan <gshan@redhat.com>

Thanks,
Gavin