On 12.06.23 00:16, Alexandre Belloni wrote: > On 11/06/2023 18:28:22+0200, Jan Kiszka wrote: >> On 11.06.23 17:11, Alexandre Belloni wrote: >>> On 11/06/2023 15:38:04+0200, Jan Kiszka wrote: >>>> On 10.06.23 10:31, Alexandre Belloni wrote: >>>>> Hello Jan, >>>>> >>>>> On 09/06/2023 23:04:12+0200, Jan Kiszka wrote: >>>>>> From: Jan Kiszka <jan.kiszka@xxxxxxxxxxx> >>>>>> >>>>>> The VL bit in the seconds register remains set only until seconds are >>>>>> written under main power. As this often happens during boot-up after >>>>>> picking up a network time, make sure to preserve the low battery state >>>>>> across this, caching it and returning it via the RTC_VL_BACKUP_LOW bit. >>>>>> >>>>>> To permit userspace clearing this state during runtime, also implement >>>>>> RTC_VL_CLR that works against the cached state. >>>>>> >>>>>> This is emulating RTCs which have a battery voltage check that works >>>>>> under main power as well. >>>>>> >>>>> >>>>> Emulating doesn't work well and I deliberately chose to not implement >>>>> it. For example, in your scenario, if you boot twice without using >>>>> VL_READ, you anyway have lost the information. This makes emulating >>>>> unreliabl. The fix you need is in userspace where you have to ensure you >>>>> read the status before setting the time. >>>> >>>> Then let's make sure the bit is also set in the hardware register. Then >>>> also the reboot issue (which is practically a minor one) is solved. The >>>> current situation is far from optimal. >>> >>> This doesn't work because then the time will be considered invalid. I'm >>> not sure why you don't want to fix your userspace. >>> >> >> Nope, that could be easily avoided in software. The actual problem is >> that the VL bit is not settable (clear-on-write). And that means we >> can't do anything about losing the low battery information across >> reboots - but that's no difference to the situation with the existing >> driver. >> >> There is no "fix" for userspace as there is no standard framework to >> read-out the status early and retrieve it from there when the user asks >> for it. That's best done in the kernel. > > That's not true, nothing prevents userspace from reading the battery > status before setting the time and destroying the information which is > exactly what you should be doing. What is your "userspace"? Mine is stock Debian with systemd and timesyncd enabled. But there is no framework to read the status early enough and propagate that after timesyncd did its job. Any concrete suggestion to "fix" userspace? > >> >> In that light, I still believe my patch is an improvement over the >> current situation without making anything worse. > > The information goes from behaving deterministically to being unreliable > which makes the situation worse. Nope, not at all. You already lose the VL bit today during reboot when you have written a new value (which is standard). So this here is not making things worse. It's rather improving the situation for the first boot at least. Deterministically. Jan -- Siemens AG, Technology Competence Center Embedded Linux
