On 12/26/17 at 08:26am, Bhupesh Sharma wrote: > On Tue, Dec 26, 2017 at 7:58 AM, AKASHI Takahiro > <takahiro.akashi at linaro.org> wrote: > > On Tue, Dec 26, 2017 at 09:35:17AM +0800, Dave Young wrote: > >> [snip] > >> > > > Well, we may be able to change pr_warn() to pr_warn_once() here, but > >> > > > I hope that adding "numa=off" to kernel command line should also work. > >> > > > >> > > Hmm, adding "numa=off" to crashkernel bootargs works, and TBH it was > >> > > my initial thought process as well, but I am not sure if this will > >> > > cause any regressions on aarch64 systems which use crashdump feature. > >> > > >> > It should be fine since we use numa=off by default for all other arches > >> > ie. x86, ppc64 and s390. Actually disabling numa in kdump kernel can save > >> > mm component memory usage. > >> > > >> > >> Forgot to say I means in RHEL and Fedora we use numa=off for kdump.. > > > > Thank you for the clarification. > > (It might be better to make numa off automatically if maxcpus == 0 (and 1?).) > > > > Not sure if we can leave this to the distribution-specific kdump > scripts (as the crashkernel boot can be held up for sufficient time > and may appear stuck). The distribution scripts may be different (for > e.g. ubuntu and RHEL/fedora) across distributions and may have > different bootarg options. Personally I think distribution should take care of this param as for kdump. But as AKASHI said it could be a issue for 1st kernel with nr_cpus=1 booting. Problem is why we do not see this issue on other machines. > > So how about considering a kernel fix only which doesn't require > relying on changing the distribution-specific kdump scripts, as we > should avoid introducing a regression while trying to fix a regression > :) > > Just my 2 cents. > > Thanks, > Bhupesh Thanks Dave
