On Tuesday, July 21, 2020 1:20:14 AM CEST Francisco Jerez wrote: > [cut] > > If there is a bug, then what exactly is it, from the users' perspective? > > > > It can be reproduced easily as follows: > > | echo 1 > /sys/devices/system/cpu/intel_pstate/hwp_dynamic_boost > | for p in /sys/devices/system/cpu/cpufreq/policy*/energy_performance_preference; do echo performance > $p; done > > Let's make sure that the EPP updates landed on the turbostat output: > > |[..] > | Core CPU Avg_MHz Busy% Bzy_MHz HWP_REQ > | - - 1 0.05 2396 0x0000000000000000 > | 0 0 1 0.05 2153 0x0000000000002704 > | 0 4 1 0.04 2062 0x0000000000002704 > | 1 1 1 0.02 2938 0x0000000000002704 > | 1 5 2 0.09 2609 0x0000000000002704 > | 2 2 1 0.04 1857 0x0000000000002704 > | 2 6 1 0.05 2561 0x0000000000002704 > | 3 3 0 0.01 1883 0x0000000000002704 > | 3 7 2 0.07 2703 0x0000000000002704 > |[..] > > Now let's do some non-trivial IO activity in order to trigger HWP > dynamic boost, and watch while random CPUs start losing their EPP > setting requested via sysfs: > > |[..] > | Core CPU Avg_MHz Busy% Bzy_MHz HWP_REQ > | - - 16 0.81 2023 0x0000000000000000 > | 0 0 7 0.66 1069 0x0000000080002704 > ^^ > | 0 4 24 2.19 1116 0x0000000080002704 > ^^ > | 1 1 18 0.68 2618 0x0000000000002704 > | 1 5 1 0.03 2005 0x0000000000002704 > | 2 2 2 0.07 2512 0x0000000000002704 > | 2 6 33 1.35 2402 0x0000000000002704 > | 3 3 1 0.04 2470 0x0000000000002704 > | 3 7 45 1.42 3185 0x0000000080002704 > ^^ Actually, that's because intel_pstate_hwp_boost_up() and intel_pstate_hwp_boost_down() use the hwp_req_cached value for updating the HWP Request MSR and that is only written to by intel_pstate_hwp_set() which is only invoked on policy changes, so the MSR writes from intel_pstate_set_energy_pref_index() basically get discarded. So this is a matter of synchronizing intel_pstate_set_policy() with intel_pstate_set_energy_pref_index() and they both acquire intel_pstate_limits_lock already, so this shouldn't be too difficult to fix. Let me cut a patch for that.
