Package: kdump-tools Version: 1:1.6.3-2 Kernel: 4.9.0-0.bpo.5-amd64 #1 SMP Debian 4.9.65-3+deb9u2~bpo8+1 (2017-01-05) x86_64 GNU/Linux Lib: 2.19-18+deb8u10
Perhaps related to #877250 The package adds this value to kdump-tools.grub.default: GRUB_CMDLINE_LINUX_DEFAULT="$GRUB_CMDLINE_LINUX_DEFAULT crashkernel=384M-:128M" This setting may work for most boxes, but we have kubernetes hosts with large memory footprints that need at least 256M. We set this value in Puppet, but the package value is added after it in the cmdline and therefore trumps it: BOOT_IMAGE=/vmlinuz-4.9.0-0.bpo.5-amd64 <settings> crashkernel=256M <more settings> crashkernel=384M-:128M So, when these boxes crash, they get stuck (with no logging indicating what the issue is) and don’t come back without manual intervention. They also don’t record a crash dump. I chased this bad behavior for 3+ months and would really like to see it resolved at the package level. My suggestion is to remove the value or replace it with crashkernel=auto where supported. Also, better logging (something like “Unable to create crash dump. Crashkernel size too small.”) would have helped me solve this issue in a day. Thanks! Alice
