Re: [PATCH RESEND] alarmtimer: Use aie_timer from RTC device instead of own timer

[Thomas Gleixner]

On Tue, Oct 01 2024 at 08:41, Mario Limonciello wrote:
> On 10/1/2024 03:30, Thomas Gleixner wrote:
>> On Mon, Sep 30 2024 at 13:29, Mario Limonciello wrote:
>>> It was reported that suspend-then-hibernate stopped working with modern
>>> systemd versions on an AMD Cezanne system. The reason for this breakage
>>> was because systemd switched to using alarmtimer instead of the wakealarm
>>> sysfs file.
>>>
>>> The wakealarm sysfs file programs the time to the `rtc->aie_timer` member
>>> of the RTC, whereas the alarmtimer suspend routine programs it to it's
>>> own device.
>>>
>>> On AMD Cezanne systems rtc_read_alarm() is used to program a secondary
>>> timer with the value of the timer. This behavior was introduced by
>>> commit 59348401ebed9 ("platform/x86: amd-pmc: Add special handling
>>> for timer based S0i3 wakeup").
>>>
>>> As rtc_read_alarm() uses the `rtc->aie_timer` to report the cached
>>> timer no alarm is provided as enabled.
>>>
>>> To fix this issue, drop the use of a dedicated timer for the alarmtimer
>>> and instead use `rtc->aie_timer` in the alarmtimer suspend/resume
>>> routines.
>> 
>> I'm not sure that this is correct. There is a reason why alarmtimer uses
>> a dedicated timer 
>
> Do you know what it is?  When I was looking at this problem I wasn't sure.

Because you cannot just blindly overwrite the aie_timer. It might have
been set by something else. Both end up in the timerqueue and the first
expiring timer is armed first.

>> and this worked correctly so far.
>> 
>> I'd rather look at commit 59348401ebed9, which plays games with the RTC.
>
> The workaround in commit 59348401ebed9 exists because of what appears to 
> be a platform bug that is unique to Cezanne systems.  Newer stuff 
> (Mendocino, Rembrandt, Phoenix, Strix etc) doesn't need or use it.

The problem is that this hack looks at the aie_timer and not at the
first expiring timer in the timerqueue which arms the RTC. That's
obviously wrong. There is no interface to get that information, but
that's a solvable problem.

Thanks,

        tglx