[Bug 1728115] Re: crashkernel offset prevents kernel boot

[Łukasz Zemczak]

I don't see this fix in bionic yet. Could anyone first release it there?
Stable updates can only be backported if they're present in the devel
series.

-- 
You received this bug notification because you are a member of Ubuntu
Sponsors Team, which is subscribed to the bug report.
https://bugs.launchpad.net/bugs/1728115

Title:
  crashkernel offset prevents kernel boot

Status in linux package in Ubuntu:
  Incomplete
Status in makedumpfile package in Ubuntu:
  In Progress
Status in linux source package in Artful:
  New
Status in makedumpfile source package in Artful:
  In Progress
Status in linux source package in Bionic:
  Incomplete
Status in makedumpfile source package in Bionic:
  In Progress

Bug description:
  [Impact]
  On Power Systems, the kernel won't boot when given a crashkernel parameter with an offset/start address of 32M. No output will be shown, giving no clue to the user why the system has not booted, or what the problem is.

  [Test Case]
  Installing kdump-tools on artful, then booting the system. It won't boot. With the fix, it boots and the crash kernel is reserved.

  [Regression Potential]
  Some Power Systems might have problems loading the kernel at this address. LP#1567539 is not really clear if PowerNV systems won't kdump when using an address different from 32M. However, it has been requested from an IBM person to test it with 128M instead, and no particular problem was shown. It's possible that there was no reason at first to use 32M, and no problems will be found on either PowerNV or other systems on the field. On the other hand, it's possible they might break kdump. But, right now, those systems won't even boot the first kernel without this change.

  ----

  The linux kernel won't boot when crashkernel parameter tells it to
  load a crash kernel at 32Mi on ppc64el on artful.

  This happens because the artful kernel is too big. In fact, multiple
  requirements on the architecture lead to that:

  Kernel memory at address 0 is reserved.
  crashkernel must be at first RMO, so architecture puts it at 128Mi. However, kdump-tools currently puts it at @32Mi because of bug https://bugs.launchpad.net/ubuntu/+source/makedumpfile/+bug/1567539.
  PACA and LPPACA need to be at the first RMO as well, and with 2048 CPUs, they take more than 5MB and 2MiB, respectively.

  With the kernel now taking around 25MB from stext to _end, the kernel
  can't reserve enough memory for PACA or LPPACA right after boot, and
  it panics.

  So, right after installing kdump-tools on artful, and rebooting, the
  kernel won't boot, with no sign of life as we haven't even started any
  console. Investigation for this issue took an entire day.

  The fix would be setting the loading address to 128MiB, and start
  reducing size of PACA and maybe remove some of the requirements for
  the location of PACA and crash kernel.

  I would not even set the loading address of the crash kernel in the
  parameter itself, and leave it to the kernel to decide it, which it
  already does and already would put it at 128Mi.

  Cascardo.

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1728115/+subscriptions