Hi, On Wed, May 10, 2023 at 9:42 AM Doug Anderson <dianders@xxxxxxxxxxxx> wrote: > > Hi, > > On Wed, May 10, 2023 at 9:30 AM Mark Rutland <mark.rutland@xxxxxxx> wrote: > > > > On Wed, May 10, 2023 at 08:28:17AM -0700, Doug Anderson wrote: > > > Hi, > > > > Hi Doug, > > > > > On Wed, Apr 19, 2023 at 3:57 PM Douglas Anderson <dianders@xxxxxxxxxxxx> wrote: > > > > This is an attempt to resurrect Sumit's old patch series [1] that > > > > allowed us to use the arm64 pseudo-NMI to get backtraces of CPUs and > > > > also to round up CPUs in kdb/kgdb. The last post from Sumit that I > > > > could find was v7, so I called this series v8. I haven't copied all of > > > > his old changelongs here, but you can find them from the link. > > > > > > > > Since v7, I have: > > > > * Addressed the small amount of feedback that was there for v7. > > > > * Rebased. > > > > * Added a new patch that prevents us from spamming the logs with idle > > > > tasks. > > > > * Added an extra patch to gracefully fall back to regular IPIs if > > > > pseudo-NMIs aren't there. > > > > > > > > Since there appear to be a few different patches series related to > > > > being able to use NMIs to get stack traces of crashed systems, let me > > > > try to organize them to the best of my understanding: > > > > > > > > a) This series. On its own, a) will (among other things) enable stack > > > > traces of all running processes with the soft lockup detector if > > > > you've enabled the sysctl "kernel.softlockup_all_cpu_backtrace". On > > > > its own, a) doesn't give a hard lockup detector. > > > > > > > > b) A different recently-posted series [2] that adds a hard lockup > > > > detector based on perf. On its own, b) gives a stack crawl of the > > > > locked up CPU but no stack crawls of other CPUs (even if they're > > > > locked too). Together with a) + b) we get everything (full lockup > > > > detect, full ability to get stack crawls). > > > > > > > > c) The old Android "buddy" hard lockup detector [3] that I'm > > > > considering trying to upstream. If b) lands then I believe c) would > > > > be redundant (at least for arm64). c) on its own is really only > > > > useful on arm64 for platforms that can print CPU_DBGPCSR somehow > > > > (see [4]). a) + c) is roughly as good as a) + b). > > > > > It's been 3 weeks and I haven't heard a peep on this series. That > > > means nobody has any objections and it's all good to land, right? > > > Right? :-P > > > > FWIW, there are still longstanding soundness issues in the arm64 pseudo-NMI > > support (and fixing that requires an overhaul of our DAIF / IRQ flag > > management, which I've been chipping away at for a number of releases), so I > > hadn't looked at this in detail yet because the foundations are still somewhat > > dodgy. > > > > I appreciate that this has been around for a while, and it's on my queue to > > look at. > > Ah, thanks for the heads up! We've been thinking about turning this on > in production in ChromeOS because it will help us track down a whole > class of field-generated crash reports that are otherwise opaque to > us. It sounds as if maybe that's not a good idea quite yet? Do you > have any idea of how much farther along this needs to go? I'm still very interested in this topic. Do you have anything concrete that is broken? Your reply gives me the vibe that there have been a bunch of bugs found recently and, while there are no known issues, you're worried that there might be something lingering still. Is that correct, or do you have something concrete that's broken? Is this anything others could help with? I could make an attempt, or we might be able to rope some of the others interested in this topic and more GIC-knowledgeable to help? I still have a goal for turning this on for production but obviously don't want to make the device crashier because of it. Also: if this really has known bugs, should we put a big WARN_ON splat if anyone tries to turn on pseudo NMIs? Without that, it feels like it's a bit of a footgun. > ...of > course, we've also run into issues with Mediatek devices because they > don't save/restore GICR registers properly [1]. In theory, we might be > able to work around that in the kernel. To followup with this, we've formulated a plan for dealing with Mediatek Chromebooks. I doubt anyone is truly interested, but if anyone is the details are in the public Google bug tracker [1]. None of that should block lading the NMI backtrace series. > In any case, even if there are bugs that would prevent turning this on > for production, it still seems like we could still land this series. > It simply wouldn't do anything until someone turned on pseudo NMIs, > which wouldn't happen till the kinks are worked out. > > ...actually, I guess I should say that if all the patches of the > current series do land then it actually _would_ still do something, > even without pseudo-NMI. Assuming the last patch looks OK, it would at > least start falling back to using regular IPIs to do backtraces. That > wouldn't get backtraces on hard locked up CPUs but it would be better > than what we have today where we don't get any backtraces. This would > get arm64 on par with arm32... The more I thought about it, the more I liked the idea of going full hog on the idea in patch #10. I've posted v9 [2] where I've really embraced the idea of falling back to a regular IPI if NMI isn't available. Hopefully it looks keen. [1] https://issuetracker.google.com/281831288 [2] https://lore.kernel.org/r/20230601213440.2488667-1-dianders@xxxxxxxxxxxx -Doug
