Re: [RFC v1 0/8] acpi/x86: s2idle: Introduce and implement runtime standby ABI for ACPI s0ix platforms
From: Mario Limonciello
Date: Fri Apr 24 2026 - 09:48:30 EST
On 4/23/26 20:35, Dmitry Osipenko wrote:
On 4/22/26 11:58, Antheas Kapenekakis wrote:
On Wed, 22 Apr 2026 at 10:51, Antheas Kapenekakis <lkml@xxxxxxxxxxx> wrote:
On Wed, 22 Apr 2026 at 02:56, Dmitry Osipenko
<dmitry.osipenko@xxxxxxxxxxxxx> wrote:
Hi, Antheas
On 3/17/26 14:04, Dmitry Osipenko wrote:
On 3/16/26 22:36, Antheas Kapenekakis wrote:
On Mon, 16 Mar 2026 at 20:33, Antheas Kapenekakis <lkml@xxxxxxxxxxx> wrote:
On Mon, 16 Mar 2026 at 20:03, Dmitry Osipenko
<dmitry.osipenko@xxxxxxxxxxxxx> wrote:
Hi Antheas,
On 1/15/26 10:49, Antheas Kapenekakis wrote:
On Thu, 15 Jan 2026 at 01:07, Dmitry Osipenko
<dmitry.osipenko@xxxxxxxxxxxxx> wrote:
On 1/13/26 13:11, Antheas Kapenekakis wrote:
<snip>
Hi Dmitry,
let me go inline.
The primary goal is to support screen-off DSM for a power-efficient
background games downloading [1] and further resume-to-dark on Steam
Deck and other handhelds. There is no strict timeline, usual "sooner the
better". Downstreams will use customized WIP solution till upstream will
get necessary generic interfaces.
[1] https://store.steampowered.com/news/app/1675200/view/771930569635267984
Ok, this makes things clearer. I had done some testing to see the
viability of such approach.
One big problem [1] had was that the compression algorithm that Steam
used was very CPU intensive. However, it was announced that that
changed, which makes low power downloads more viable.
However, even so, I do not think the sleep DSM is designed for
prolonged background use and certain devices might overheat.
Specifically, I think the Go S disables its fan while in that DSM.
Looking back to what Windows does, it only uses the Sleep state to do
periodic polling, and if there are updates it transitions to display
off.
This is a fair approach for [1]. For example, device wakes up every
two hours while connected to a charger, stays on sleep state, checks
for updates, and if there are any and conditions are met, transitions
to display off and starts downloading.
However, this means you do not get a smaller tdp limit. Given you
control the unfrozen userspace in that state though, such a limit does
not help either. The device will use what it needs to for downloads.
This makes the SD 5W low power mode puzzling, as it means downloads
will potentially take longer and I would be punished as a user for
using that mode. Instead, Steam should be optimized to use less than
5W or perhaps 10W when downloading from gigabit in some way.
Two more considerations in this case are that a lot of devices will
turn off their controllers when entering display off. And the rest
when entering sleep. This is good because when you are in dark resume,
the RGB of the device has turned off. But for [1] it is problematic
because it assumes the controller works and is what is used to wake
the device so the mode is broken. For Legion, Sleep is used to turn
off the controller, and for other devices Sleep Entry/Exit. New in ROG
Xbox Ally devices is that the controller no longer turns off, but it
is muted.
The other consideration is that three additional patches are needed
for ROG Ally devices to work correctly with this series, 2 cleanup
commits and 1 small delay. But after that it should be drop in. I
cannot comment on the new hid drivers for Asus and Legion that are
currently being developed. Particularly, hid-legion-go(?) has a
reset_resume() cb where it should have used resume? Or not anything?
The legion controllers save os mode until they disconnect, which they
do with this series, so the driver would always re-initialize on
wake-up.
My rough understanding that a firmware/BIOS update may be needed for
some devices to leverage DSM in regards to power consumption
improvement. Could be true that practically it may not improve much,
will see. Even if not all current devices will benefit from the
screen-off DSM, it may differ for a later generations.
Hi Dmitry,
sorry, I have been busy the past few days. I read through Rafael's
comments, I will respond to them in the next 2-3 days.
Your understanding is correct. You will see 0 performance difference
with the screen off DSM. It is not supposed to affect the thermal
envelope. Turning off the screens of your laptop due to inactivity is
not supposed to affect the thermal envelope.
For devices that use the DSM to turn off their controller (~60%),
there might be a marginal <.5W improvement. The rest use the sleep
DSMs for that. Those do affect the thermal envelope. However, they are
not designed for prolonged CPU intensive tasks such as a SteamOS
download mode. This is the bigger concern. From my experience, I
expect the Go S to overheat.
The improvement because of turning off the controllers / RGB is
something that is needed for download mode in any case. And as my
previous comment suggests, lowering the thermal limit may not be
required / advantageous.
Here, the sleep dsms are also important, because of their broad
effects to power light / thermal envelope. As they can be used for
brief wake / update checks. For battery powered devices, perhaps these
are not important (excl. hibernation checks), but for desktop consoles
it is expected functionality.
A common approach for upstreaming is to divide problem into smaller
manageable parts. That's what I'm planning to focus on now to see if we
can start easy with a minimal changes.
Sure. One potential approach for this is this series, where the first
part does the plumbing and the second part the exposing. They can be
merged independently.
I also made sure to address Rafael's comments, so the ABI of this
series is completely independent of ACPI, S0ix or whether the device
has a display. I also removed all references to Intel, AMD specific
power envelope terminology. Moreover, most of the logic now resides in
suspend.c and the hooks are in platform_ calls, so it can be
implemented for other platforms easily.
However, the first part of this series does some refactorings which
assume a favorable outcome. If we do not want to assume that, a
simpler initial series would just move the MS/display on/off DSMs to
.begin() in s2idle.c. If you think that would be easier to merge, you
are welcome to start with that. Then this series would be refactored
on top and merged as a single unit. Keep in mind the ROG Ally conflict
would also arise in this case as well.
Please don't worry about the credit. You did a significant ground work
that is well recognized by now. Thanks a lot for your efforts and help.
Starting from scratch of course won't be a good approach with all the
broad testing you've done.
Great. Sounds good to me.
I'm taking latest version of your patches and will update them in
accordance to the review from Rafael.
Rafael raised some good questions I need to respond to. Specifically,
the ABI is not yet decided, so the comments are not ready to address
yet.
Nonetheless, this series is ready to test with the current ABI.
Moreover, Rafael suggests for upstreaming to follow essentially what I
outlined. An initial series should move the DSMs to the beginning of
the suspend sequence, and the follow-up would implement the ABI.
I was planning to work on the first sub-series as a non-RFC after
responding to Rafael's comments. Let me know how you plan to proceed.
Please go on and make a v2. This version works well with my testing. Thanks!
Do you have updates on v2? For now I don't have good suggestion on where
to put and how to define standby knobs in sysfs in a more generic
manner. Pick what feels better or keep current /sys/power variant with
the rest of review comments addressed.
Please let me know if help needed,
Hi Dmitry,
I had some work pop up the past three weeks so I could not go ahead
with this. I will reprioritize and get back to you with an update by
end of week.
Here is the current plan for V2. It will consist of four parts.
i) We keep the initial rename patches.
ii) Then, we move all the calls to the begin() and end() functions.
After these, we perform a summary cleanup of asus-wmi and hid-asus to
undo the duplicate handling of the calls due to the prior sequence
causing issues on Ally devices, and add a small 500ms quirk after the
begin calls for these devices to resolve a timing issue present on all
firmwares (the Ally units have a hardcoded ~300ms delay in which they
fade their RGB before disconnecting the controller and saving its
state, causing it to reboot on resume if that does not happen; this is
common in all aura devices such as Asus Keyboards, but no other
devices cause issues due to the state loss).
These first two parts can merge initially to ensure no issues arise.
Then, we implement the activity API in two parts.
iii) Introduce drivers/acpi/activity.c (match platform profile for
now), which has an ABI of
If this ends up being the agreeable direction I don't see a reason to structure it in drivers/acpi initially in the first place.
/sys/class/activity/[driver_name]/{name,hint,hint_choices} (non-ACPI
specific). The ABI and activity.c do not perform state transitions,
match the platform_profile internal API structure, and are not ACPI
specific (the file can be moved later).
I note here the use of [driver_name] instead of platform-profile-N,
which should allow basic userspace apps to skip the loop and name
check, hardcoding /sys/class/activity/s2idle/{hint,hint_choices}. In
addition, prefixing the files with hint allows reusing the scaffolding
to implement other ABIs in the future, which would be hard to do with
platform-profile.
Doesn't this imply the driver_name becomes part of the ABI itself? Couldn't userspace just read /sys/class/activity/0/name?
The whole point of stable interfaces is to try to come up with something that would scale even for situations that don't make sense at the time.
An unfortunate example we should avoid repeating is RAPL.
It's named "intel-rapl" but there are MSRs on AMD SoC's that provide the same functionality. Maybe at the time it made sense to keep 'intel' in the name as it was originally an Intel interface, but now any SoC that provides that kind of interface has that in the path name.
For WMI drivers, [driver-name] would include an integer suffix as per
convention for these drivers, with the first instance using 0.
iv) Register an activity handler in s2idle.c and push the transition
logic in it. This way, other users do not have to abide by it. If
userspace is not activity hint aware, enter the state snooze in
begin(), and restore the prior state in end(). If there is a timing
quirk (such as for the Ally), we will store the jiffies of the last
transition and compare them in begin(). This way, when userspace
becomes activity hint-aware and enters snooze prior to beginning the
suspend sequence, the delay is shortened. Since userspace freeze by
the systemd freezer group (if enabled) and then suspend.c require
around 100ms-300ms, this eats more than half of the delay in this case
for traditional resume / suspend, and eliminates it for dark screen
wakeups that remain in snooze.
Thanks for the update. Will see what others will say about the activity
interface.