intel_pstate fails to enable ITMT if CPPC v2 is disabled by firmware, resulting in bad single-core performance on some heterogeneous-core Intel systems

Bug #2069988 reported by Aaron Rainbolt
6
This bug affects 1 person
Affects Status Importance Assigned to Milestone
linux (Ubuntu)
New
Undecided
Unassigned

Bug Description

This issue is explained and discussed in detail on the linux-acpi mailing list in the following three threads:

* Initial issue report: https://<email address hidden>/T/#u
* First patch submitted to mainline for review: https://<email address hidden>/T/#t
* Second patch (not mine) submitted to mainline for review: https://<email address hidden>/T/#t

Intel-based machines using heterogeneous-core processors (i.e., anything with P-cores and E-cores) require Intel Turbo Boost Max Technology 3.0 (ITMT) in order for the EEVDF scheduler to correctly assign intensive single-core loads to P-cores. If ITMT is not enabled, the scheduler does not get hints about CPU performance differences, causing it to oftentimes schedule single-core workloads on E-cores rather than P-cores. This leads to a serious drop in single-core performance.

Currently, if CPPC v2 is not acked by the firmware, the intel_pstate driver will *not* enable ITMT. This means that any device with a hybrid-core Intel CPU, where CPPC v2 is missing, non-functional, or incorrectly disabled by the firmware, will experience the performance drop mentioned above. This happens in the wild:

* My workplace, Kubuntu Focus, is affected by this issue: https://<email address hidden>/T/#u
* A user of a Star Labs laptop was also affected: https://github.com/StarLabsLtd/firmware/issues/143
* Someone on the Manjaro forums appears to have run into this potentially: https://forum.manjaro.org/t/linux-kernel-6-6-lts-cpu-regression-on-i7-alderlake/157474
* Someone onf the Gentoo forums also may have run into this: https://forums.gentoo.org/viewtopic-p-8819389.html?sid=5997f89fd5a202b6db8396fba0b45821

This issue, as shown above, has been discussed on the linux-acpi mailing list, and currently has a fix in the works. https://<email address hidden>/T/#t The fix works by falling back to using info from MSR_HWP_CAPABILITIES in the event CPPC is not available.

To post a comment you must log in.
This report contains Public information  
Everyone can see this information.

Other bug subscribers

Remote bug watches

Bug watches keep track of this bug in other bug trackers.