On Fri, Oct 1, 2021 at 5:00 AM Guillaume Tucker <guillaume.tucker@xxxxxxxxxxxxx> wrote: > > Hello Rob, I'm adding a few more here as I think there's a wider topic of probing and modules. > > On 28/07/2021 11:51, Heiko Stübner wrote: > > Am Mittwoch, 28. Juli 2021, 11:16:14 CEST schrieb Marc Zyngier: > >> On Wed, 28 Jul 2021 09:59:49 +0100, > >> Guillaume Tucker <guillaume.tucker@xxxxxxxxxxxxx> wrote: > >>> > >>> On 28/07/2021 09:39, Robin Murphy wrote: > >>>> Hi Guillaume, > >>>> > >>>> Not sure what I did to get CC'd on this, but since I'm here... > >>> > >>> You were listed by get_maintainer.pl for the patch found by the > >>> bisection: > >>> > >>> Robin Murphy <robin.murphy@xxxxxxx> (authored:1/8=12%,added_lines:9/71=13%,removed_lines:16/41=39%,added_lines:11/45=24%,removed_lines:18/32=56%,authored:1/12=8%,added_lines:22/83=27%,removed_lines:29/69=42%) > >>> > >>> Maybe the logic to automatically build the list of recipients > >>> could look at those stats and apply some threshold if too many > >>> people get listed because of small contributions to some files. > >>> It's not a common issue though, usually the recipients are all > >>> pretty relevant. > >>> > >>>> On 2021-07-28 07:04, Guillaume Tucker wrote: > >>>>> Please see the bisection report below about usb2phy failing to > >>>>> probe on rk3399-gru-kevin. > >>>>> > >>>>> Reports aren't automatically sent to the public while we're > >>>>> trialing new bisection features on kernelci.org but this one > >>>>> looks valid. > >>>>> > >>>>> The bisection was run in the Renesas tree but the same regression > >>>>> is present in mainline for both usb2phy0 and usb2phy1 devices: > >>>>> > >>>>> https://linux.kernelci.org/test/plan/id/6100af012344eef9b85018f3/ > >>>>> https://linux.kernelci.org/test/case/id/6100af012344eef9b85018fa/ > > This issue is still present and it got bisected yet again > yesterday by KernelCI. > > >>>>> I don't see any errors in the logs, it looks like the driver is > >>>>> just not probing. > >>>> > >>>> What's the actual testcase for "rockchip-usb2phy0-probed"? If it's looking for a hard-coded path like "/sys/bus/platform/devices/ff770000.syscon:usb2-phy@e450/driver" then it can be expected to fail, since changing the node name is reflected in the device name. /sys/bus/platform/devices/ paths are not an ABI. I'll consider nodenames an ABI if a change is noticed, but not for sysfs path. >From sysfs testings/sys-devices (what /sys/bus/platform/devices/ links to): What: /sys/devices Date: February 2006 Contact: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> Description: The /sys/devices tree contains a snapshot of the internal state of the kernel device tree. Devices will be added and removed dynamically as the machine runs, and between different kernel versions, the layout of the devices within this tree will change. Please do not rely on the format of this tree because of this. If a program wishes to find different things in the tree, please use the /sys/class structure and rely on the symlinks there to point to the proper location within the /sys/devices tree of the individual devices. Or rely on the uevent messages to notify programs of devices being added and removed from this tree to find the location of those devices. Note that sometimes not all devices along the directory chain will have emitted uevent messages, so userspace programs must be able to handle such occurrences. > >>> > >>> Dang, you're right. This is the test case: > >>> > >>> https://github.com/kernelci/bootrr/blob/main/boards/google%2Ckevin#L119 > >>> > >>> assert_driver_present rockchip-usb2phy-driver-present rockchip-usb2phy > >>> assert_device_present rockchip-usb2phy0-probed rockchip-usb2phy ff770000.syscon:usb2-phy@e450 > >>> assert_device_present rockchip-usb2phy1-probed rockchip-usb2phy ff770000.syscon:usb2-phy@e460 Why do you even care about the USB PHY probing directly? It is not usable on its own. What you care about is whether the USB controller(s) probed and USB is working. > >>> Now that needs a conditional depending on the kernel version. Or > >>> we could try to make it more dynamic rather than with hard-coded > >>> paths, but doing that has its own set of issues too. > >> > >> And this shows once more that DT churn has consequences: it breaks a > >> userspace ABI. Changing userspace visible paths for the sake of > >> keeping a build-time checker quiet seems counter-productive. My > >> preference would be to just revert this patch, and instead have an > >> annotation acknowledging the deviation from the 'standard' and keeping > >> the checker at bay. > > > > I'd be fine with that, if that is the consensus. And an annotation comment > > would be good in that case, just to keep a similar change from getting > > submitted. > > > > I guess the interesting question is if dtbscheck has some sort of tooling > > to detect these "this is meant to be that way for backwards compatibility" > > hence adding Rob for that question. I would like to have some way to disable specific warnings and/or have warning levels. The problem is there's not any sort of identifier to key off of at a granularity smaller than a schema file. So in this case, we'd have to disable the PHY binding entirely (that's not actually too bad here). The second problem is where do we put that control. We could do some sort of source annotation. That would have to be maintained in the yaml encoded output and also would not work when we don't have dts source. The schema checks currently require dts source as we use some of the annotations like /bits/ that are lost in the dtb, but addressing that to support running the checks on a running system is something I'm actively working on. > Could you please take a look at Heiko's suggestion above to see > if this should be solved in dtbs_check? If not then we would > need to change the KernelCI test definition to look for a > different name based on the kernel version (which sounds like > breaking user-space). Looking at the above check, that looks horrible to create and maintain (regardless of this issue). I was actually just wondering if there were any tests for devices that didn't probe. On the recent VExpress breakages, we only noticed on the boards that failed to init their timer. Newer boards that use the arch timer still booted, but a bunch of devices didn't probe. Can't we extract every device not bound to a driver, get the compatible string or modalias, and then compare that to the aliases of all the modules? Of course, that doesn't work in the built-in case. That's perhaps something we should fix though. (I'd also like to be able to extract all driver DT compatibles at build time and have the same issue.) Alternatively, couldn't this test list the compatible string instead of node name and then you search devices for a match? Or maybe there is a much more simple fix. We could just log failed probes in a standard way for tools to understand. Rob
