In my presentation at AsiaBSDCon, I proposed moving this functionality from user space into the kernel. One limitation of the current design is that when a task migrates from one CPU core to another, powerd must effectively relearn or re-evaluate the workload characteristics before making frequency decisions. The kernel, however, already has direct visibility into workload scheduling and task migration events. By keeping the policy logic in the kernel, workload information can be carried across core migrations, enabling more responsive and efficient CPU frequency management.
For the interface, cpufreq driver exposes min and max performance value (can be frequency or an abstracted value like cppc provided) it can provides. The kernel detects the work load and set the desired performance between min and max performance.
On 6/3/26 02:25, Adrian Chadd wrote:
.. it sounds like we want to split this up into package versus cluster frequency or something? What's the API look like right now? -adrian On Mon, 1 Jun 2026 at 23:42, ShengYi Hung <[email protected]> wrote:The branch main has been updated by aokblast: URL: https://cgit.FreeBSD.org/src/commit/?id=7b26353a59d66dc1bc611fd042a49b9e3bd13699 commit 7b26353a59d66dc1bc611fd042a49b9e3bd13699 Author: ShengYi Hung <[email protected]> AuthorDate: 2026-06-01 09:46:37 +0000 Commit: ShengYi Hung <[email protected]> CommitDate: 2026-06-02 06:41:41 +0000 hwpstate_intel: Disable package control on hybrid CPU In package control mode, the performance of all cores depends on the most recent value written to the request field. If the last write comes from an E-core, all cores are forced to align with the E-core performance level, resulting in significant performance degradation. Therefore, package control is disabled on hybrid-core systems. Reviewed by: olce MFC after: 2 weeks Sponsored by: The FreeBSD Foundation Sponsored by: Framework Computer Inc Differential Revision: https://reviews.freebsd.org/D57377 --- sys/x86/cpufreq/hwpstate_intel.c | 21 +++++++++++++++++++++ 1 file changed, 21 insertions(+) diff --git a/sys/x86/cpufreq/hwpstate_intel.c b/sys/x86/cpufreq/hwpstate_intel.c index 3d2cc0a5966b..db8600d7b89a 100644 --- a/sys/x86/cpufreq/hwpstate_intel.c +++ b/sys/x86/cpufreq/hwpstate_intel.c @@ -321,9 +321,19 @@ out: return (ret); } +static void +intel_hwpstate_hybrid_cb(void *ctx) +{ + uint32_t *small_cores = ctx; + + atomic_add_32(small_cores, PCPU_GET(small_core)); +} + void intel_hwpstate_identify(driver_t *driver, device_t parent) { + uint32_t small_cores = 0; + if (device_find_child(parent, "hwpstate_intel", DEVICE_UNIT_ANY) != NULL) return; @@ -343,6 +353,17 @@ intel_hwpstate_identify(driver_t *driver, device_t parent) if ((cpu_power_eax & CPUTPM1_HWP) == 0) return; + /* + * On hybrid-core systems, package-level control cannot be used. + * It may cause all cores to run at the E-core frequency because + * the resulting package frequency depends on the last core that + * sets the frequency. + */ + smp_rendezvous_cpus(all_cpus, smp_no_rendezvous_barrier, + intel_hwpstate_hybrid_cb, smp_no_rendezvous_barrier, &small_cores); + if (small_cores > 0 && small_cores < mp_ncores) + hwpstate_pkg_ctrl_enable = false; + if (BUS_ADD_CHILD(parent, 10, "hwpstate_intel", device_get_unit(parent)) == NULL) device_printf(parent, "hwpstate_intel: add child failed\n");
OpenPGP_0x5F4ED7D812606C60.asc
Description: OpenPGP public key
OpenPGP_signature.asc
Description: OpenPGP digital signature
