On 05/26/2015 12:32 PM, Ashwin Chaugule wrote:
On 26 May 2015 at 13:23, Al Stone al.stone@linaro.org wrote:
On 05/25/2015 03:49 PM, Ashwin Chaugule wrote:
CPPC:
CPPC (Collaborative Processor Performance Control) is a new way to control CPU performance using an abstract continous scale as against a discretized P-state scale which is tied to CPU frequency only. It is defined in the ACPI 5.0+ spec. In brief, the basic operation involves:
OS makes a CPU performance request. (Can provide min and max tolerable bounds)
Platform (such as BMC) is free to optimize request within requested bounds depending
on power/thermal budgets etc.
- Platform conveys its decision back to OS
The communication between OS and platform occurs through another medium called (PCC) Platform communication Channel. This is a generic mailbox like mechanism which includes doorbell semantics to indicate register updates. See drivers/mailbox/pcc.c
This patchset introduces a CPPC based CPUFreq driver that works with existing governors such as ondemand. The CPPC table parsing and the CPPC communication semantics are abstracted into separate files to allow future CPPC based drivers to implement their own governors if required.
Initial patchsets included an adaptation of the PID governor from intel_pstate.c. However recent experiments led to extensive modifications of the algorithm to calculate CPU busyness. Until it is verified that these changes are worthwhile, the existing governors should provide for a good enough starting point for ARM64 servers.
Finer details about the PCC and CPPC spec are available in the latest ACPI 5.1 specification.[2]
Changes since V4:
- Misc cleanups. Addressed feedback from Rafael.
- Made acpi_processor.c independent of C-states, P-states and others.
- Per CPU scanning for _CPC is now made from acpi_processor.c
- Added new Kconfig options for legacy C states and P states to enable future
support for newer alternatives as defined in the ACPI spec 6.0.
Changes since V3:
- Split CPPC backend methods into separate files.
- Add frontend driver which plugs into existing CPUfreq governors.
- Simplify PCC driver by moving communication space mapping and read/write into client drivers.
Changes since V2:
- Select driver if !X86, since intel_pstate will use HWP extensions instead.
- Added more comments.
- Added Freq domain awareness and PSD parsing.
Changes since V1:
- Create a new driver based on Dirks suggestion.
- Fold in CPPC backend hooks into main driver.
Changes since V0: [1]
- Split intel_pstate.c into a generic PID governor and platform specific backend.
- Add CPPC accessors as PID backend.
[1] - http://lwn.net/Articles/608715/ [2] - http://www.uefi.org/sites/default/files/resources/ACPI_5_1release.pdf [3] - https://patches.linaro.org/40705/
Ashwin Chaugule (6): PCC: Initialize PCC Mailbox earlier at boot ACPI: Make ACPI processor driver more extensible ACPI: Introduce CPU performance controls using CPPC CPPC: Add a CPUFreq driver for use with CPPC CPPC: Probe for CPPC tables for each ACPI Processor object PCC: Enable PCC only when needed
drivers/acpi/Kconfig | 58 ++- drivers/acpi/Makefile | 8 +- drivers/acpi/cppc_acpi.c | 808 ++++++++++++++++++++++++++++++++++++++++ drivers/acpi/processor_driver.c | 89 +++-- drivers/cpufreq/Kconfig | 2 +- drivers/cpufreq/Kconfig.arm | 16 + drivers/cpufreq/Kconfig.x86 | 2 + drivers/cpufreq/Makefile | 2 + drivers/cpufreq/cppc_cpufreq.c | 197 ++++++++++ drivers/mailbox/Kconfig | 2 +- drivers/mailbox/pcc.c | 2 +- include/acpi/cppc_acpi.h | 137 +++++++ include/acpi/processor.h | 118 +++++- 13 files changed, 1380 insertions(+), 61 deletions(-) create mode 100644 drivers/acpi/cppc_acpi.c create mode 100644 drivers/cpufreq/cppc_cpufreq.c create mode 100644 include/acpi/cppc_acpi.h
Apart from how the patches showed up in email :), nice work, Ashwin.
Can you add a description of how you tested this, too?
This was tested on an SBSA compatible ARMv8 server with CPPCv2 firmware running on a remote processor. I verified that each CPUs performance limits were detected and that new performance requests were made by the on-demand governor proportional to the load on each CPU. I also verified that using the acpi_processor driver correctly maps the physical CPU ids to logical CPU ids, which helps in picking up the proper _CPC details from a processor object, in the case where CPU physical ids may not be contiguous.
Excellent. Thanks, exactly what I was looking for.