Powering off idle CPUs saves about 33 mW compared to using WFI only.
Additional power savings are possible by idling the L2 and downclocking
the cluster when all CPUs are idle.
Entry and exit latency were measured using a logic analyzer, with GPIO
pins toggled in Linux after the calls to trace_cpu_idle() in
cpuidle_enter_state(), and in the power management firmware after CPU
power-off completes and immediately after detecting an interrupt.
800 us and 1500 us are worst-case values, largely driven by the fact
that the power management firmware is single threaded. It can only
handle commands to power off CPUs one at a time, and it cannot process
any commands while powering on a CPU in response to an interrupt.
The cluster suspend process reliably takes 36 us; I rounded this up to
50 us. If all CPUs enter the cluster idle state at the same time, exit
latency is actually reduced, because there is no contention in that
case. However, if only some CPUs enter the cluster idle state, behavior
is the same as for CPU idle.
Polling delay for the power management firmware to detect a pending
interrupt is insignificant; it is less than 20 us.
min-residency was chosen as the point where enabling the idle state
consumed no more average power than disabling the idle state at a
variety of interrupt rates.
Signed-off-by: Samuel Holland <samuel@xxxxxxxxxxxx>
---
I'm sending this patch as an RFC because it raises questions about how
we handle firmware versioning. How far back does (or should) our support
for old TF-A and Crust versions go?
cpuidle has a problem that without working firmware support, CPUs will
enter idle states and be unable to wake up. As a result, the system will
hang at some point during boot, usually before getting to userspace.
For over a year[0], TF-A has exposed the PSCI CPU_SUSPEND function when
a SCPI implementation is present[1]. Implementing CPU_SUSPEND is
required for implementing SYSTEM_SUSPEND[2], even if CPU_SUSPEND is not
itself used for anything.
However, there was no code to actually wake up a CPU once it called the
CPU_SUSPEND function, because I could not find the register providing
the necessary information. The fact that CPU_SUSPEND was broken affected
nobody, because nothing ever called it -- there were no idle states in
the DTS. In hindsight, what I should have done was always return failure
from sunxi_validate_power_state(), but that ship has long sailed.
I finally found the elusive register and implemented the wakeup code
earlier this month[3]. So now, CPU_SUSPEND actually works, if all of
your firmware is up to date, and cpuidle works if you add the states in
your device tree.
Unfortunately, there is currently nothing verifying that compatibility.
So you can get into four possible scenarios:
1) No idle states in DTS, any firmware => Linux works, with baseline
power consumption.
2) Idle states added to DTS, no Crust/SCPI => Linux works, but every
attempt to enter an idle state is rejected because CPU_SUSPEND is
not hooked up. So power consumption increases by a sizable amount.
3) Idle states added to DTS, "old" Crust/SCPI (before [3]) => Linux
fails to boot, because CPUs never return from idle states.
4) Idle states added to DTS, "new" Crust/SCPI (after [3]) => Linux
works, with improved power consumption compared to the baseline.
Obviously, we want to prevent scenario 3 if possible.
Enter the current patch: I chose the arm,psci-suspend-param values
specifically so they would be _rejected_ by the current TF-A code. This
makes scenario 3 behave like scenario 2. I then have some follow-up TF-A
patches (not yet submitted) to switch to the new parameter encoding[4].
This brings me back to my original question. Once the TF-A patches in
[4] are merged, scenario 3 (with an updated TF-A but an old Crust) would
fail to boot again. Do we care?
Should I implement some kind of runtime version checking, so TF-A can
disable CPU_SUSPEND if it would be broken? Or instead, should we wait
some amount of time to merge this patch (or the patches at [4]) and
assume people have upgraded?
Where would people expect this sort of possibly-breaking change to be
documented?
Separately, since I assume most A64/H5 users (outside of LibreELEC and
the PinePhone) are not using Crust, scenario 2 would be very common. If
merging this patch increases their idle power draw by 500 mW, is that an
acceptable cost for decreasing other users' idle power draw by 50 mW?
Sorry for the wall of text,
Samuel
[0]: https://git.trustedfirmware.org/TF-A/trusted-firmware-a.git/commit/plat/allwinner/common/sunxi_pm.c?id=e382c88e2a26995099bb931d49e754dcaebc5593
[1]: https://git.trustedfirmware.org/TF-A/trusted-firmware-a.git/tree/plat/allwinner/common/sunxi_scpi_pm.c?id=2e0e51f42586826a1f6f6c1e532f90e6df642cf5#n190
[2]: https://git.trustedfirmware.org/TF-A/trusted-firmware-a.git/tree/lib/psci/psci_setup.c?id=2e0e51f42586826a1f6f6c1e532f90e6df642cf5#n251
[3]: https://github.com/crust-firmware/crust/commits/85944467c804
[4]: https://github.com/crust-firmware/arm-trusted-firmware/commits/d6ebf5dab2da
---
arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi | 26 +++++++++++++++++++
arch/arm64/boot/dts/allwinner/sun50i-h5.dtsi | 26 +++++++++++++++++++
2 files changed, 52 insertions(+)
diff --git a/arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi b/arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi
index 57786fc120c3..2b1b5b36098c 100644
--- a/arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi
+++ b/arch/arm64/boot/dts/allwinner/sun50i-a64.dtsi
@@ -54,6 +54,7 @@ cpu0: cpu@0 {
clocks = <&ccu CLK_CPUX>;
clock-names = "cpu";
#cooling-cells = <2>;
+ cpu-idle-states = <&cpu_sleep>, <&cluster_sleep>;
};
cpu1: cpu@1 {
@@ -65,6 +66,7 @@ cpu1: cpu@1 {
clocks = <&ccu CLK_CPUX>;
clock-names = "cpu";
#cooling-cells = <2>;
+ cpu-idle-states = <&cpu_sleep>, <&cluster_sleep>;
};
cpu2: cpu@2 {
@@ -76,6 +78,7 @@ cpu2: cpu@2 {
clocks = <&ccu CLK_CPUX>;
clock-names = "cpu";
#cooling-cells = <2>;
+ cpu-idle-states = <&cpu_sleep>, <&cluster_sleep>;
};
cpu3: cpu@3 {
@@ -87,6 +90,29 @@ cpu3: cpu@3 {
clocks = <&ccu CLK_CPUX>;
clock-names = "cpu";
#cooling-cells = <2>;
+ cpu-idle-states = <&cpu_sleep>, <&cluster_sleep>;
+ };
+
+ idle-states {
+ entry-method = "psci";
+
+ cpu_sleep: cpu-sleep {
+ compatible = "arm,idle-state";
+ local-timer-stop;
+ entry-latency-us = <800>;
+ exit-latency-us = <1500>;
+ min-residency-us = <25000>;
+ arm,psci-suspend-param = <0x00010003>;
+ };
+
+ cluster_sleep: cluster-sleep {
+ compatible = "arm,idle-state";
+ local-timer-stop;
+ entry-latency-us = <850>;
+ exit-latency-us = <1500>;
+ min-residency-us = <50000>;
+ arm,psci-suspend-param = <0x01010013>;
+ };
};
L2: l2-cache {
diff --git a/arch/arm64/boot/dts/allwinner/sun50i-h5.dtsi b/arch/arm64/boot/dts/allwinner/sun50i-h5.dtsi
index 578a63dedf46..1c416f648c58 100644
--- a/arch/arm64/boot/dts/allwinner/sun50i-h5.dtsi
+++ b/arch/arm64/boot/dts/allwinner/sun50i-h5.dtsi
@@ -18,6 +18,7 @@ cpu0: cpu@0 {
clocks = <&ccu CLK_CPUX>;
clock-latency-ns = <244144>; /* 8 32k periods */
#cooling-cells = <2>;
+ cpu-idle-states = <&cpu_sleep>, <&cluster_sleep>;
};
cpu1: cpu@1 {
@@ -28,6 +29,7 @@ cpu1: cpu@1 {
clocks = <&ccu CLK_CPUX>;
clock-latency-ns = <244144>; /* 8 32k periods */
#cooling-cells = <2>;
+ cpu-idle-states = <&cpu_sleep>, <&cluster_sleep>;
};
cpu2: cpu@2 {
@@ -38,6 +40,7 @@ cpu2: cpu@2 {
clocks = <&ccu CLK_CPUX>;
clock-latency-ns = <244144>; /* 8 32k periods */
#cooling-cells = <2>;
+ cpu-idle-states = <&cpu_sleep>, <&cluster_sleep>;
};
cpu3: cpu@3 {
@@ -48,6 +51,29 @@ cpu3: cpu@3 {
clocks = <&ccu CLK_CPUX>;
clock-latency-ns = <244144>; /* 8 32k periods */
#cooling-cells = <2>;
+ cpu-idle-states = <&cpu_sleep>, <&cluster_sleep>;
+ };
+
+ idle-states {
+ entry-method = "psci";
+
+ cpu_sleep: cpu-sleep {
+ compatible = "arm,idle-state";
+ local-timer-stop;
+ entry-latency-us = <800>;
+ exit-latency-us = <1500>;
+ min-residency-us = <25000>;
+ arm,psci-suspend-param = <0x00010003>;
+ };
+
+ cluster_sleep: cluster-sleep {
+ compatible = "arm,idle-state";
+ local-timer-stop;
+ entry-latency-us = <850>;
+ exit-latency-us = <1500>;
+ min-residency-us = <50000>;
+ arm,psci-suspend-param = <0x01010013>;
+ };
};
};
--
2.26.2
