On Mon, Oct 05, 2020 at 02:06:37PM +0100, Marc Zyngier wrote: > On 2020-10-05 12:22, Thierry Reding wrote: > > On Mon, Oct 05, 2020 at 12:14:40PM +0100, Marc Zyngier wrote: > > > Jon recently reported that one of the Tegra systems (Jetson TX2, aka > > > tegra186) stopped booting with the introduction of the "IPI as IRQs" > > > series. After a few weeks of head scratching and complete puzzlement, > > > I obtained a board and started looking at what was happening. > > > > > > The interrupt hierarchy looks like this: > > > > > > [DEVICE] -A-> [PMC] -B-> [GIC] > > > > > > which seems simple enough. However, not all the devices attached to > > > the PMC follow this hierarchy, and in some cases, the 'B' link isn't > > > present in the HW. In other cases, neither 'A' nor 'B' are present. > > > And yet the PMC driver creates such linkages using random hwirq values > > > for the non-existent links, potentially overriding existing mappings > > > in the process. "What could possibly go wrong?" > > > > Yes, that would've been my fault. It seemed like the right thing to do > > at the time, but the way you describe it makes it obvious that it was > > not. I can't say I understand why this would've worked prior to the > > rework that made this surface, though. > > Because until these IPI patches, the range 0-7 never ever appeared > as actual hwirqs in the GIC domain. SGIs were handled in the GIC > code, behind the core kernel's back. As soon as we start using > an actual domain mapping for hwirq 0, the PMC driver starts messing > with it. Okay, that makes sense. > > > It turns out that for the 'B' link, the PMC driver uses hwirq 0, which > > > is SGI0 for the GIC, and used as the rescheduling IPI. Obviously, this > > > doesn't go very well, nor very far, as the IPI gets routed to random > > > drivers. Also, as the handling flow has been overridden, this > > > interrupt never gets deactivated and can't fire anymore. Yes, this is > > > bad. > > > > > > The 'A' link is less problematic, but the hwirq value is still out of > > > the irqdomain range, and gets remapped every time a new 'A'-less > > > driver comes up. > > > > > > Instead, let's trim the unused hierarchy levels as needed. This > > > requires some checks in the upper levels of the hierarchy as we now > > > have optional levels, but this looks a lot saner than what we > > > currently have. With this, tegra186 is back booting on -next. > > > > > > I haven't tested any wake-up stuff, nor any other nvidia system (this > > > is the only one I have). If people agree to these changes, I can take > > > them via the irqchip tree so that they make it into the next merge > > > window. > > > > Yeah, it sounds like this needs to go in ideally before the rework that > > caused this to surface in order to preserve bisectibility. But if it > > goes in afterwards that's probably fine as well. > > It's easy to take it as part of the same pull request as the IPI > stuff. Not fully bisectable for this platform, but close enough. > I may even be able to merge this in *before* the IPI patches > (I'd need to rebuild irq/irqchip-next, but that won't alter the commit > ids of the individual patches as they are on separate branches). I noticed there's a small conflict with another patch that I've queued up and that I was thinking of sending in for v5.10. This adds a callback for .irq_retrigger. I don't exactly recall why I needed this, but I've been carrying that patch locally for a while now. I don't think it's critical, but there must've been an issue that it fixed. Unfortunately I seem to have been in a hurry because the commit message is a bit terse. I can probably just leave that out for now to not further complicated things, in which case it should be fine to take your patches here into v5.10 along with the IPI changes. As I already mentioned I'm a bit unhappy about the needless move of the irq_chip to a global variable, but I can send out an updated version with those changed back, which is basically what I have on my local test branch because I had to manually apply that patch anyway due to the conflict I mentioned above, so keeping the embedded irq_chip was actually simpler and reduces the diff a little bit. > > Let Jon and myself do a bit of testing with this to verify that the wake > > up paths are still working. > > Sure. Let me know what you find. The results are in and it's a bit of a mixed bag. I was able to confirm that Tegra194 also boots again after this series and I'm also able to resume from sleep using either rtcwake or the power-key as wakeup source, so the wake-events mechanism is still functional after the series. I do see a bit of breakage on resume, but none of that seems related to your patches and is likely something that crept in while we were looking into the current issue. Jon had started a job in our test farm in parallel and that came back with a failing suspend/resume test on Tegra186 (Jetson TX2), but that seems to have been a pre-existing issue. This was already in linux-next around next-20200910 and Jon had been investigating it when the boot failures due to the IPI changes started happening. So I then hooked up my Jetson TX2 and verified locally that I can properly suspend/resume using either rtcwake or the power-key as wakeup source, just like I previously did on Tegra194 (Jetson AGX Xavier). Tegra186 seems to be a little more unstable because it didn't boot every time for me, but that is probably not related to this. I also did a bit of testing on Tegra210 and that seems to also still work. We cannot test suspend/resume and therefore wake-events there out-of-the-box because of some long-standing issue with USB refusing to suspend, but I was able to at least trigger wake-up using the RTC by first unloading the XUSB driver. It doesn't resume to a fully functional system, but if I do see it resume, then the wake-event code must have worked. So, I'm tempted to say: Tested-by: Thierry Reding <treding@xxxxxxxxxx> on the series and then we can pick up the other pieces once everything is back to at least booting. Thanks again for tracking this down, and apologies for breaking it in the first place. Let me know if you want me to send the bike-shedded version of your patches. Thierry
Attachment:
signature.asc
Description: PGP signature
