Re: [PATCH] kdump: add default crashkernel reserve kernel config options

[Dave Young]

On 05/21/18 at 12:02pm, Andrew Morton wrote:
> On Mon, 21 May 2018 10:53:37 +0800 Dave Young <dyoung@xxxxxxxxxx> wrote:
> 
> > This is a rework of the crashkernel=auto patches back to 2009 although
> > I'm not sure if below is the last version of the old effort:
> > https://lkml.org/lkml/2009/8/12/61
> > https://lwn.net/Articles/345344/
> > 
> > I changed the original design, instead of adding the auto reserve logic
> > in code, in this patch just introduce two kernel config options for
> > the default crashkernel value in MB and the threshold of system memory
> > in MB so that only reserve default when system memory is equal or
> > above the threshold.
> > 
> > With the kernel configs distributions can easily change the default
> > values so that people do not need to manually set kernel cmdline
> > for common use cases and one can still overwrite the default value
> > with manual setup or disable it by using crashkernel=0
> > 
> > Signed-off-by: Dave Young <dyoung@xxxxxxxxxx>
> > ---
> > Another difference is with original design the crashkernel size scales
> > with system memory, according to test, large machine may need more
> > memory in kdump kernel because of several factors:
> > 1. cpu numbers, because of the percpu memory allocated for cpus.
> >    (kdump can use nr_cpus=1 to workaround this, but some
> >     arches do not support nr_cpus=X for example powerpc) 
> > 2. IO devices, large system can have a lot of io devices, although we
> >    can try to only add those device drivers we needed, it is still a
> >    problem because of some built-in drivers, some stacked logical devices
> >    eg. device mapper devices, acpi etc.  Even if only considering the
> >    meta data for driver model it will still be a big number eg. sysfs
> >    files etc.
> > 3. The minimum memory requirement for some device drivers are big, even
> >    if some of them have implemented low meory profile.  It is usual to see
> >    10M memory use for a storage driver.
> > 4. user space initramfs size growing.  Busybox is not usable if we need
> >    to add udev support and some complicate storage support.  Use dracut
> >    with systemd, especially networking stuff need more memory.
> > 
> > So probably add another kernel config option to scale the memory size
> > eg.  CRASHKERNEL_DEFAULT_SCALE_RATIO is also good to have,  in RHEL we
> > use base_value + system_mem >> (2^14) for x86.  I'm still hesatating
> > how to describe and add this option. Any suggestions will be appreciated.
> > 
> > ...
> >
> > --- linux-x86.orig/arch/Kconfig
> > +++ linux-x86/arch/Kconfig
> > @@ -10,6 +10,22 @@ config KEXEC_CORE
> >  	select CRASH_CORE
> >  	bool
> >  
> > +config CRASHKERNEL_DEFAULT_THRESHOLD_MB
> > +	int "System memory size threshold for kdump memory default reserving"
> > +	depends on CRASH_CORE
> > +	default 0
> > +	help
> > +	  CRASHKERNEL_DEFAULT_MB is used as default crashkernel value if
> > +	  the system memory size is equal or bigger than the threshold.
> 
> "the threshold" is rather vague.  Can it be clarified?
> 
> In fact I'm really struggling to understand the logic here....

Sorry I missed this comment, the threshold is like this:

if system memory size is lower than the threshold then we will do
nothing and do not reserve crashkernel memory by default.  Eg. if the
threshold is 1024M then default reservation is only used when system has
more than 1024M memory, and for those low mem machine we do not reserve by
default.

Thanks
Dave