On Tuesday, November 26, 2019 4:17:09 PM CET Leonard Crestez wrote:
Support for frequency limits in dev_pm_qos was removed when cpufreq was switched to freq_qos, this series attempts to restore it by reimplementing on top of freq_qos.
Discussion about removal is here: https://lore.kernel.org/linux-pm/VI1PR04MB7023DF47D046AEADB4E051EBEE680@VI1P...
The cpufreq core switched away because it needs contraints at the level of a "cpufreq_policy" which cover multiple cpus so dev_pm_qos coupling to struct device was not useful. Cpufreq could only use dev_pm_qos by implementing an additional layer of aggregation anyway.
However in the devfreq subsystem scaling is always performed on a per-device basis so dev_pm_qos is a very good match. Support for dev_pm_qos in devfreq core is here (latest version, no dependencies outside this series):
https://patchwork.kernel.org/cover/11252409/
That series is RFC mostly because it needs these PM core patches. Earlier versions got entangled in some locking cleanups but those are not strictly necessary to get dev_pm_qos functionality.
In theory if freq_qos is extended to handle conflicting min/max values then this sharing would be valuable. Right now freq_qos just ties two unrelated pm_qos aggregations for min and max freq.
This is implemented by embeding a freq_qos_request inside dev_pm_qos_request: the data field was already an union in order to deal with flag requests.
The internal freq_qos_apply is exported so that it can be called from dev_pm_qos apply_constraints.
The dev_pm_qos_constraints_destroy function has no obvious equivalent in freq_qos and the whole approach of "removing requests" is somewhat dubios: request objects should be owned by consumers and the list of qos requests will most likely be empty when the target device is deleted. Series follows current pattern for dev_pm_qos.
First two patches can be applied separately.
Changes since v3:
- Fix s/QOS/QoS in patch 2 title
- Improves comments in kunit test
- Fix assertions after freq_qos_remove_request
- Remove (c) from NXP copyright header
- Wrap long lines in qos.c to be under 80 chars. This fixes checkpatch but the
rule is already broken by code in the files.
- Collect reviews
Link to v3: https://patchwork.kernel.org/cover/11260627/
Changes since v2:
- #define PM_QOS_MAX_FREQUENCY_DEFAULT_VALUE FREQ_QOS_MAX_DEFAULT_VALUE
- #define FREQ_QOS_MAX_DEFAULT_VALUE S32_MAX (in new patch)
- Add initial kunit test for freq_qos, validating the MAX_DEFAULT_VALUE found
by Matthias and another recent fix. Testing this should be easier! Link to v2: https://patchwork.kernel.org/cover/11250413/
Changes since v1:
- Don't rename or EXPORT_SYMBOL_GPL the freq_qos_apply function; just
drop the static marker. Link to v1: https://patchwork.kernel.org/cover/11212887/
Leonard Crestez (4): PM / QoS: Initial kunit test PM / QoS: Redefine FREQ_QOS_MAX_DEFAULT_VALUE to S32_MAX PM / QoS: Reorder pm_qos/freq_qos/dev_pm_qos structs PM / QoS: Restore DEV_PM_QOS_MIN/MAX_FREQUENCY
drivers/base/Kconfig | 4 ++ drivers/base/power/Makefile | 1 + drivers/base/power/qos-test.c | 117 ++++++++++++++++++++++++++++++++++ drivers/base/power/qos.c | 73 +++++++++++++++++++-- include/linux/pm_qos.h | 86 ++++++++++++++----------- kernel/power/qos.c | 4 +- 6 files changed, 242 insertions(+), 43 deletions(-) create mode 100644 drivers/base/power/qos-test.c
I have applied the whole series as 5.5 material, but I have reordered the fix (patch [2/4]) before the rest of it and marked it for -stable.
Thanks!