Hm - sounds related. I updated the GZ to 20160411T120144Z, but the docker container still exhibits the behaviour (unless there are other updates I need to apply?)
On Mon, Apr 11, 2016 at 2:44 PM, Jerry Jelinek <[email protected]> wrote: > It is possible that you are hitting https://smartos.org/bugview/OS-5311, > which I just fixed on Friday. To summarize the problem: > > The bug is that the active contract template is only setup on the initial > LWP in the process, so any time that LWP forks, the child will be in a > different contract, as configured. However, if any other LWP initiates the > fork, then that LWP has a NULL active contract template, so the child will > be in the same contract. > > Due to the how we setup the application process for a Docker zone, if the > application forks a child, it must be in a different contract. If it's not, > then when the child exits, it will cause all of the other processes in the > same contract to die and the zone will halt. > > Jerry > > > On Mon, Apr 11, 2016 at 6:49 AM, Nigel Magnay <[email protected]> > wrote: > >> Ok, so I have traced a little further >> >> Part of the maven build shells out to run git (via the plugin code at >> [2]; the last lines of log show this before the mysterious death: >> >> [INFO] Executing: /bin/sh -c cd >> '/home/jenkins/realtime/commons/csw-commons-utils' && 'git' 'rev-parse' >> '--verify' 'HEAD' >> [INFO] Working directory: /home/jenkins/realtime/commons/csw-commons-utils >> [INFO] Storing buildNumber: a20f2793de8f449ff5c638479e73dab4a884936c at >> timestamp: 1460376264112 >> >> The output is showing that the git process returned (the info output >> showing the returned info). However, in the dtrace log, I can see this: >> >> 937/1: lx_emulate(7fffff08eca0, 231, [8d, 0, 8d, ffffffffffffff90, 3c, >> e7]) >> 937/1: lx_exit_common(LX_ET_EXIT_GROUP, 141) >> >> Which I'm assuming is the git process returning an error from exit(), and >> googling around 141 seems to indicate a pipe being closed (e.g: [1]) >> >> Sometimes the build does not die at that point - and the corresponding >> part of the log did not show 141, but instead a 0 >> >> 1264/1: lx_exit_common(LX_ET_EXIT_GROUP, 0) >> >> If I remove the part of the build that calls git, it completes >> successfully, so it certainly seems to be related to this. >> >> What I don't really understand is why the child process exiting is >> causing the parent to mysteriously exit - but not immediately ? >> >> >> [1] https://bugs.launchpad.net/mksh/+bug/1532621 >> [2] >> https://github.com/mojohaus/buildnumber-maven-plugin/blob/buildnumber-maven-plugin-1.3/src/main/java/org/codehaus/mojo/build/CreateMojo.java >> >> >> >> >> >> On Mon, Apr 11, 2016 at 10:37 AM, Nigel Magnay <[email protected]> >> wrote: >> >>> So uname of the GZ is >>> SunOS headnode 5.11 joyent_20160317T000105Z i86pc i386 i86pc >>> >>> The docker container is >>> Linux ff4a058f3f5e 3.13.0 BrandZ virtual linux x86_64 GNU/Linux >>> >>> I've attached the Dockerfile, though it's not that exciting. I seem to >>> get failures quite often when the docker api chooses the HN to provision a >>> node rather than a CN, but I'm not totally sure about that. I don't think >>> it's running out of memory - vmstat in the GZ shows >>> >>> I've attached an output from the tail of truss on the GZ for the java >>> process (bear in mind I know nothing currently about dtrace, but I will go >>> away and try to read up) - I assume though that I'll see segfaults as a >>> natural product of the java memory management. >>> >>> The only thing that stands out to my eye in the trace is >>> vforkx(0) = 65788 >>> >>> which seems an strange return value if I'm reading it correctly? >>> >>> >>> On Sat, Apr 9, 2016 at 11:16 AM, Nigel Magnay <[email protected]> >>> wrote: >>> >>>> I'll dig out the Dockerfile on monday - it's not particularly complex. >>>> >>>> Basically the (jenkins) build process asks triton to instantiate a >>>> (java:8) docker image, the uses ssh to connect and invoke maven. I can >>>> connect 'by hand' and also make it die - though it may be that it's only >>>> doing this on some nodes. >>>> >>>> The most likely explanation is I've either misunderstood or >>>> misconfigured something. There would seem to be enough free memory and swap >>>> on the host whilst it's running, but might there be other limits >>>> (processes?) that may have been breached causing it to be killed? >>>> >>>> >>>> >>>> On Fri, Apr 8, 2016 at 10:20 PM, Elijah Zupancic <[email protected]> >>>> wrote: >>>> >>>>> Hi Nigel, >>>>> >>>>> Could you please give us some additional information. Could you tell >>>>> us the platform image version. You can find it by doing a uname -a or if >>>>> you are within a docker container a /native/bin/uname -a >>>>> >>>>> Also, with regards to the Docker build process are you using Docker >>>>> build with Triton or are you using Docker build with Linux Docker and then >>>>> running the built image on Triton? >>>>> >>>>> Also, if you Dockerfile isn't sensitive could you please share it with >>>>> us? I can attempt to reproduce. Another way to narrow down problems is to >>>>> try to run it in the Joyent public cloud and to see if it dies there. >>>>> >>>>> Thanks, >>>>> Elijah >>>>> >>>>> On Fri, Apr 8, 2016 at 6:11 AM, Nigel Magnay <[email protected]> >>>>> wrote: >>>>> >>>>>> Hi - >>>>>> >>>>>> I've successfully built a Triton cloud, and am using it to provision >>>>>> docker containers to build our java software. >>>>>> >>>>>> The java container I use is derived from "java:8" docker image. >>>>>> >>>>>> Mostly it works. However I am getting mysterious failures where the >>>>>> build process (maven) simply exits half way through with no error. I >>>>>> have a >>>>>> test system that is spun up that, mostly, consistently fails in the same >>>>>> place. >>>>>> >>>>>> dmesg from inside the container yields nothing. Irritatinly, if I do >>>>>> 'strace -ff -o me ...', the build continues past the point it was >>>>>> failing. >>>>>> If I try to monitor the process from a second connection - sometimes the >>>>>> build continues and *that* connection just dies. >>>>>> >>>>>> I was able to connect briefly with truss from the GZ, but I'm not >>>>>> sure it tells me very much. >>>>>> >>>>>> >>>>>> Could I have hit some process accounting limit? I don't think it's >>>>>> memory (the process is set to a fairly low java Xmx limit). Where else >>>>>> can >>>>>> I look to figure this out? >>>>>> >>>>>> >>>>> >>>>> >>>>> -- >>>>> -Elijah >>>>> >>>> >>>> >>> >> > *smartos-discuss* | Archives > <https://www.listbox.com/member/archive/184463/=now> > <https://www.listbox.com/member/archive/rss/184463/22541698-24d6dc34> | > Modify > <https://www.listbox.com/member/?&;> > Your Subscription <http://www.listbox.com> > ------------------------------------------- smartos-discuss Archives: https://www.listbox.com/member/archive/184463/=now RSS Feed: https://www.listbox.com/member/archive/rss/184463/25769125-55cfbc00 Modify Your Subscription: https://www.listbox.com/member/?member_id=25769125&id_secret=25769125-7688e9fb Powered by Listbox: http://www.listbox.com
