On Wednesday 22 February 2017 11:24:13 Hans de Goede wrote: > HI, > > On 22-02-17 09:49, Pali Rohár wrote: > >On Wednesday 22 February 2017 09:36:08 Hans de Goede wrote: > >>Hi, > >> > >>On 21-02-17 18:08, Pali Rohár wrote: > >>>On Tuesday 21 February 2017 17:14:06 Hans de Goede wrote: > >>>>Hi, > >>>> > >>>>On 21-02-17 16:13, Pali Rohár wrote: > >>>>>On Tuesday 21 February 2017 15:56:43 Hans de Goede wrote: > >>>>>>>So do we really need this code which prevents update? > >>>>>> > >>>>>>Yes, because the ABI specification for the new brightness_hw_changed says > >>>>>>that poll() listeners will only be woken up if the brightness is changed > >>>>>>outside of the kernel and not when the kernel changes it itself. > >>>>> > >>>>>So in case there are two applications in userspace which want to monitor > >>>>>brightness change and both of those application could change brightness > >>>>>(via sysfs) then these two applications would not know about every > >>>>>brightness change and would be out-of-sync of reality what is really > >>>>>configured by kernel. > >>>> > >>>>Yes, because with triggers and blinking etc. it is impossible for > >>>>userspace to continuously monitor brigthness in a way which does not > >>>>cause a high system load. > >>> > >>>Triggers and blinking features are out due to high cpu load. Fine. > >>> > >>>But why also manual writes to /sys/class/leds/... by userspace > >>>applications is filtered and not reported via poll()? > >> > >>I agree that having a way for interested userspace to detect those > >>would be good, but that would need to be another API, may poll() > >>on the brightness attribute itself while excluding triggers / blinking > >>changes from wakeup ? > >> > >>Anyways that is something to discuss in a thread specific to the > >>LED subsystem and somewhat orthogonal to this patch. > > > >Ok, lets start discussion about it in new separate thread. I was in > >impression that this was already part of discussion and in proposed ABI. > >Now I see that it was just in original cpu consuming ABI which was > >rejected. > > > >>One disadvantage of poll() is that it does not give the source of > >>the change, so in retrospect I actually like the new brightness_hw_changed > >>attribute as that does give the source, which is something which we need > >>to know in userspace. > > > >Do you really in userspace need to know source of change? And to > >distinguish between change from hardware and change from userspace done > >by echo > /sys/class/leds? > > Yes, a change done through hardware (through the firmware handled > hotkeys) will make the desktop environment show an osd overlay > with the new kbd backlight brightness, where as a change done > through e.g. the brightness slider in gnome-control-panel should > not show that same osd. > > >I though that this is for informing userspace application that led > >status was changed and application should update some bar or number > >which show current state of backlight level. > > That is only part of it, we also want the OSD but ONLY when changed > through the hotkeys (on other models the hotkeys are handled in > userspace software, which will then also show the OSD, we want this > to be consistent whether the keys are handled in firmware or in > userspace). Ok, so userspace application wants to distinguish between these types of events: 1) autonomous hardware decided to change brightness 2) application itself changed brightness 3) other application changed brightness 4) pressed keyboard hotkey changed brightness 4.1) managed by firmware 4.2) managed by some application And for 4) and 1) you want to show notification to user, right? > >>In previous versions of the ABI I had to do > >>the same brightness comparison I'm doing in the dell-laptop driver > >>now in userspace, where it can never be done safely as userspace does > >>not know about other userspace. > >> > >>Since the Dell smbios events don't provide us with a source of the > >>change, we need to compare the brightness to the last set brightness > >>somewhere and IMHO the kernel is the best (least bad) place to do > >>that. > > > >Maybe kernel place is the least bad place, but still it is unreliable > >for Dell machines. > > > >You do not know latency and delay how long it will take Dell firmware to > >deliver event to kernel. It also implies that there is race condition > >between delivering event and reading new backlight level. > > A race condition which will always exists if userspace polls the backlight > periodically say once a minute then it can always end up polling just > before the hotkey press is done. Which is exactly why we need the event, > so we don't need to poll and when the event is delivered we know the new > backlight level. So instead userspace polling you will ask for backlight level after receiving firmware event. This is same race as in userspace. Returned backlight level does not have to be one which caused firmware event. If firmware delays that event (and due to ACPI slowness it can be really delayed) then another backlight change could happen... And you will just use incorrect level for comparison. Basically on Dell machine you cannot distinguish between events done by 1), 2), 3) and 4). You are trying to hide this fact by bringing some logic which on first look can fix it... Also calling another SMBIOS call via SMM for every received event (even if userspace is not interested in it) does not look like good implementation. And another one is issued after userspace receive that event (via poll()) and try to know current value. -- Pali Rohár pali.rohar@xxxxxxxxx
