Hi, On Wed, Sep 20, 2023 at 12:26 AM Johan Hovold <johan@xxxxxxxxxx> wrote: > > On Tue, Sep 19, 2023 at 11:15:46AM -0700, Doug Anderson wrote: > > On Tue, Sep 19, 2023 at 12:07 AM Johan Hovold <johan@xxxxxxxxxx> wrote: > > > > But regardless of what a long-term proper solution to this may look > > > like, we need to fix the regression in 6.6-rc1 by restoring the old > > > behaviour. > > > > OK, fair enough. I'll take a look at your patch, though I think the > > person that really needs to approve it is Benjamin... > > > > Style-wise, I will say that Benjamin really wanted to keep the "panel > > follower" code out of the main probe routine. Some of my initial > > patches adding "panel follower" looked more like the results after > > your patch but Benjamin really wasn't happy until there were no > > special cases for panel-followers in the main probe routine. This is > > why the code is structured as it is. > > Ok, I prefer not hiding away things like that as it obscures what's > really going on, for example, in this case, that you register a device > without really having probed it. I can see your reasoning and I think that intuition is why the earlier versions of my patches had explicit "panel follower" logic in probe. However, Benjamin really liked the logic abstracted out. > As I alluded to in the commit message, you probably want to be able to > support second-source touchscreen panel followers as well at some point > and then deferring checking whether device is populated until the panel > is powered on is not going to work. Yeah, I've been pondering this too. I _think_ it would work OK-ish if both devices probed and then only one of the two would actually make the sub-HID devices. So you'd actually see both devices succeed at probing but only one of them would actually be functional. It's a bit ugly, though. :( Maybe marginally better would be if we could figure out how to have the device which fails to get its interrupt later unbind itself, if that's possible... The only other thought I had would be to have the parent i2c bus understand that it had children that were panel followers, which it should be able to do by seeing the "panel" attribute in their device tree. Then the i2c bus could itself register as a panel follower and could wait to probe its children until they were powered on. This could happen in the i2c core so we didn't have to add code like this to all i2c bus drivers. ...and, if necessary, we could add this to other busses like SPI. It feels a little awkward but could work. > I skimmed the thread were you added this, but I'm not sure I saw any > reason for why powering on the panel follower temporarily during probe > would not work? My first instinct says we can't do this, but let's think about it... In general the "panel follower" API is designed to give all the decision making about when to power things on and off to the panel driver, which is controlled by DRM. The reason for this is from experience I had when dealing with the Samsung ATNA33XC20 panel that's on "sc7180-trogdor-homestar". The TCON on that panel tended to die if you didn't sequence it just right. Specifically, if you were sending pixels to the panel and then stopped then you absolutely needed to power the panel off and on again. Folks I talked to even claimed that the panel was working "to spec" since, in the "Power Sequencing" section of the eDP spec it clearly shows that you _must_ turn the panel off and on again after you stop giving it bits. ...this is despite the fact that no other panel I've worked with cares. ;-) On homestar, since we didn't have the "panel follower" API, we ended up adding cost to the hardware and putting the panel and touchscreens on different power rails. However, I wanted to make sure that if we ran into a similar situation in the future (or maybe if we were trying to make hardware work that we didn't have control over) that we could solve it. The other reason for giving full control to the panel driver is just how userspace usually works. Right now userspace tends to power off panels if they're not used (like if a lid is closed on a laptop) but doesn't necessarily power off the touchscreen. Thus if the touchscreen has the ability to keep things powered on then we'd never get to a low power state. The above all explains why panel followers like the touchscreen shouldn't be able to keep power on. However, you are specifically suggesting that we just turn the power on temporarily during probe. As I think about that, it might be possible? I guess you'd have to temporarily block DRM from changing the state of the panel while the touchscreen is probing. Then if the panel was off then you'd turn it on briefly, do your probe, and then turn it off again. If the panel was on then by blocking DRM you'd ensure that it stayed on. I'm not sure how palatable that would be or if there are any other tricky parts I'm not thinking about. > > Thinking that way, is there any reason you can't just move the > > i2c_hid_init_irq() into __do_i2c_hid_core_initial_power_up()? You > > could replace the call to enable_irq() with it and then remove the > > `IRQF_NO_AUTOEN` flag? I think that would also solve the issue if you > > wanted to use a 2nd source + the panel follower concept? Both devices > > would probe, but only one of them would actually grab the interrupt > > and only one of them would actually create real HID devices. We might > > need to do some work to keep from trying again at every poweron of the > > panel, but it would probably be workable? I think this would also be a > > smaller change... > > That was my first idea as well, but conceptually it is more correct to > request resources at probe time and not at some later point when you can > no longer fail probe. > > You'd also need to handle the fact that the interrupt may never have > been requested when remove() is called, which adds unnecessary > complexity. I don't think it's a lot of complexity, is it? Just an extra "if" statement... -Doug
