buildman stops (crashed) on current master

[Simon Glass]

Hi Stefano,

On Wed, 20 Oct 2021 at 03:55, Stefano Babic <sbabic at denx.de> wrote:
>
> On 20.10.21 05:42, Simon Glass wrote:
> > Hi,
> >
> > On Tue, 19 Oct 2021 at 17:01, Tom Rini <trini at konsulko.com> wrote:
> >>
> >> On Tue, Oct 19, 2021 at 04:59:15PM -0600, Simon Glass wrote:
> >>> Hi Tom,
> >>>
> >>> On Tue, 19 Oct 2021 at 16:53, Tom Rini <trini at konsulko.com> wrote:
> >>>>
> >>>> On Tue, Oct 19, 2021 at 05:39:12PM +0200, Stefano Babic wrote:
> >>>>> Hi Simon,
> >>>>>
> >>>>> On 07.10.21 15:43, Simon Glass wrote:
> >>>>>> Hi Stefano,
> >>>>>>
> >>>>>> On Thu, 7 Oct 2021 at 04:37, Stefano Babic <sbabic at denx.de> wrote:
> >>>>>>>
> >>>>>>> Hi all,
> >>>>>>>
> >>>>>>> CI stops by building aarch64 without notice, for reference:
> >>>>>>>
> >>>>>>> https://source.denx.de/u-boot/custodians/u-boot-imx/-/jobs/332319
> >>>>>>>
> >>>>>>> There is no error, just process is killed. It looks like it stops at
> >>>>>>> xilinx_zynqmp_virt,
> >>>>>>>
> >>>>>>> ./tools/buildman/buildman -o /tmp -P -E -W aarch64but board can be built
> >>>>>>> without issues.
> >>>>>>>
> >>>>>>> If I build on my host (not in docker, anyway), it generally builds fine
> >>>>>>> - but it crashes sometimes, too. On gitlab instance , it crashes.
> >>>>>>> Issue does not seem that depends on merged patches, and introduces
> >>>>>>> boards were already built successfully. Any hint ? I have also no idea
> >>>>>>> what I should look as what I see is just
> >>>>>>>
> >>>>>>> "usr/bin/bash: line 104:    24 Killed
> >>>>>>> ./tools/buildman/buildman -o /tmp -P -E -W aarch64"
> >>>>>>
> >>>>>> I cannot see that link. I am not sure what is going on. Does it say
> >>>>>> what signal killed it?
> >>>>>
> >>>>> Pipelines on our server were not public - I have enbaled now for u-boot-imx.
> >>>>>
> >>>>>>
> >>>>>> Does it sit there for an hour and timeout? If so, then I  did see that
> >>>>>> myself once recently, when the Kconfig needed stdin, but I could not
> >>>>>> quitetie it down. I think buildman would provide it, but sometimes
> >>>>>> not, apparently. So it can happen when there is an existing build
> >>>>>> there and your new one which adds Kconfig options that don't have
> >>>>>> defaults, or something like that?
> >>>>>>
> >>>>>
> >>>>> I have investigated further, and I can reproduce it on my host outside the
> >>>>> gitlab server. buildman causes a OOM, but I cannot find the cause.
> >>>>>
> >>>>> Strange enough, this happens with the "aarch64" target, and I cannot
> >>>>> reproduce it with Tom's master. So it seems that -master is ok, and somethin
> >>>>> on u-boot-imx generates the OOM.
> >>>>>
> >>>>> However....
> >>>>>
> >>>>> The OOM happens always when -2 (two boards remain) appears. I can see with
> >>>>> htop that buildman starts to allocate memory until it is exhausted (64GB RAM
> >>>>> + 8 GB swap). Then the kernel decides that it is enough and kills buildman -
> >>>>> this is what I see on Ci.
> >>>>>
> >>>>> You can see now the pipelines:
> >>>>>
> >>>>> https://source.denx.de/u-boot/custodians/u-boot-imx/-/pipelines/9520
> >>>>>
> >>>>> I have then split aarch64 and I built imx8 separately - same result. The
> >>>>> pipeline stops with xilinx board, but they have nothing to do. In fact, I
> >>>>> can build all xilinx board separately. If I run buildman -W aarch64 -x
> >>>>> xilinx, OOM is shown by another board.
> >>>>>
> >>>>> Strange enough, I can build each single board with buildman without issues,
> >>>>> neither errors nor warnongs. Just when buildman runs all together (aarch64,
> >>>>> 308 boards), the OOM is generated.
> >>>>>
> >>>>> Bisect does not help: I started bisect, and at the end this commit was
> >>>>> presented:
> >>>>>
> >>>>> commit 53a24dee86fb72ae41e7579607bafe13442616f2
> >>>>> Author: Fabio Estevam <festevam at denx.de>
> >>>>> Date:   Mon Aug 23 21:11:09 2021 -0300
> >>>>>
> >>>>>      imx8mm-cl-iot-gate: Split the defconfigs
> >>>>
> >>>> I strongly suspect what's going on here is that these new defconfigs are
> >>>> out of sync with changes now in Kconfig.  The build itself will just sit
> >>>> there, waiting for the "oldconfig" prompt to be answered.
> >>>>
> >>>> I want to say the problem here is that stdin is open, rather than
> >>>> pointing to something closed and would lead to the build failing
> >>>> immediately, rather than once a timeout is hit, or OOM kicks in due to
> >>>> kconfig chewing up all the memory.
> >>>
> >>> Yes that's exactly what I saw...
> >>>
> >>> In fact, see this commit:
> >>>
> >>> e62a24ce27a buildman: Avoid hanging when the config changes
> >>>
> >>> But that was 3 years ago.
> >>
> >> Looks like something else needs to be changed then, I've bisected down
> >> similar failures here before very recently.
> >
> > I dug into this a bit and I think buildman can detect this situation.
> > I'll send a little series.
> >
>
> Patch definetly help ;-)
>
> It breaks build (on CI when build-tools runs), but I get much more
> details when I build locally single boards. I can find for
> kontron-sl-mx8mm several errors due to:
>
> - CONFIG_SYS_LOAD_ADDR not defined in configs, but in header
> - CONFIG_SYS_EXTRA_OPTIONS instead of CONFIG_IMX_CONFIG
> - CONFIG_SYS_MALLOC_LEN not defined in config, but in header
>
> Your patch are a valueable tool (CI driove me crazy), I can now folow
> what happens. I send a patch for kontron, and I go on with the rest (I
> guess kontron is not the only board causing this deadlock). Many thanks !
>
> Tom, I apply Simon's patches on my tree, I cannot work without them...

OK good. Unfortunately this is likely to throw up problems that are
only fatal 1% of the time.

Regards,
Simon