Messages in this thread | | | Date | Thu, 7 May 2020 19:09:58 +0100 | Subject | [PATCH 00/14] Modularize schedutil | From | Quentin Perret <> |
| |
Android is trying very hard to use a single kernel image (commonly called Generic Kernel Image, or GKI), closely aligned with mainline, to run on all Android devices regardless of the vendor.
The GKI project intends to not only improve the status quo for Android users directly (less fragmentation simplifies updatability), but also to benefit upstream by forcing all vendors to agree on one common kernel, that we push hard to be aligned with mainline.
One challenge to implement GKI is to avoid bloating the kernel by compiling too many things in, especially given that different devices need different things. As such, anything that can be turned into a module helps GKI, by offering an alternative to having that component built-in. This is true for pretty much anything that can be made modular, including drivers as well as other kernel components, such as CPUFreq governors.
Indeed, in practice, Android devices often ship with only one CPUFreq governor enabled, and don't require any other that would simply waste memory for no benefits. All CPUFreq governors can already be built as modules with one notable exception: schedutil. Though popular in Android, some devices do not use schedutil, which is why it would be preferable to not have it unconditionally built in GKI. This series is an attempt to solve this problem, by making schedutil tristate.
While modularization is usually not something we want to see near the scheduler, it appeared to me as I wrote those patches that the particular case of schedutil was actually not too bad to implement. We already have to support switching governors at run-time, simply because userspace is free to do that, so the infrastructure for turning sugov on and off dynamically is already there. Loading the code a little later doesn't seem to make that a lot worse.
Patches 01-05 refactor some code to break the few dependencies on schedutil being builtin (notably EAS). Patches 06-12 export various symbols that schedutil needs when compiled as a module. And finally, patches 13-14 finish off the work by making the Kconfig tristate.
--- The series is based on Peter's sched/fifo [1] branch (because sugov uses sched_setscheduler_nocheck()).
[1] https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git/log/?h=sched/fifo
Quentin Perret (14): sched: Provide sched_set_deadline() sched: cpufreq: Use sched_set_deadline() from sugov sched: cpufreq: Introduce 'want_eas' governor flag sched: cpufreq: Move sched_cpufreq_governor_change() sched: cpufreq: Move schedutil_cpu_util() arch_topology: Export cpu_scale per-cpu array kthread: Export kthread_bind_mask() sched/core: Export runqueues per-cpu array sched/cpufreq: Export cpufreq_this_cpu_can_update() sched/fair: Export cpu_util_freq() tick/sched: Export tick_nohz_get_idle_calls_cpu x86: Export arch_scale_freq_key sched: cpufreq: Use IS_ENABLED() for schedutil sched: cpufreq: Modularize schedutil
arch/x86/kernel/smpboot.c | 1 + drivers/base/arch_topology.c | 1 + drivers/cpufreq/Kconfig | 2 +- include/linux/cpufreq.h | 6 +- include/linux/sched.h | 2 + include/linux/sched/sysctl.h | 2 +- kernel/kthread.c | 1 + kernel/sched/core.c | 18 ++++ kernel/sched/cpufreq.c | 34 ++++++ kernel/sched/cpufreq_schedutil.c | 176 +++---------------------------- kernel/sched/fair.c | 119 ++++++++++++++++++++- kernel/sched/sched.h | 36 ++----- kernel/sched/topology.c | 16 +-- kernel/sysctl.c | 2 +- kernel/time/tick-sched.c | 1 + 15 files changed, 212 insertions(+), 205 deletions(-)
-- 2.26.2.526.g744177e7f7-goog
| |