nodepool: set spec_store_bypass_disable=prctl for Jammy images
We have noticed Jammy jobs failing with DISK_FULL returns from Zuul. Upon further investigation, they have hundreds of megabytes of syslog filled with traces. As noted inline we've debugged this down to SSBD interactions with some cloud providers. Set this to the later-kernel upstream default of "prctl" to avoid this issue. Change-Id: I09a495a9b50c55b65221be56337f52388a5d9d48
This commit is contained in:
parent
10e1a87d7e
commit
9673eb9a76
|
@ -319,6 +319,25 @@ diskimages:
|
|||
env-vars:
|
||||
DIB_DISTRIBUTION_MIRROR: 'https://mirror.dfw.rax.opendev.org/ubuntu'
|
||||
DIB_DEBIAN_COMPONENTS: 'main,universe'
|
||||
# NOTE(ianw) 2022-05-20 : we have found a problem on some cloud
|
||||
# providers (OVH at least) where setting SSBD mitigation flags
|
||||
# seems to be disabled and this causes endless kernel tracebacks
|
||||
# from writing invalid flags into MSR registers, flooding the
|
||||
# logs. By default spec_store_bypass_disable is set to
|
||||
# "seccomp" which means the kernel forces every seccomp() call
|
||||
# into this. Setting it to prctl means userspace has to opt-in
|
||||
# with an explicit call. This is actually the default on more
|
||||
# recent kernels; the reasons are many but the bug [1] goes
|
||||
# through it with links to relevant changes.
|
||||
#
|
||||
# While we try and sort out either backporting this to upstream
|
||||
# kernels or some other upstream workaround, we set
|
||||
# spec_store_bypass_disable=prctl here to avoid this problem in
|
||||
# our providers. The other bits are just the dib defaults which
|
||||
# need to be replicated when overriding this.
|
||||
#
|
||||
# [1] https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1973839
|
||||
DIB_BOOTLOADER_DEFAULT_CMDLINE: 'nofb nomodeset gfxpayload=text spec_store_bypass_disable=prctl'
|
||||
|
||||
- name: ubuntu-xenial
|
||||
parent: base-debuntu
|
||||
|
|
Loading…
Reference in New Issue