[PATCH v17 02/10] x86: kdump: make the lower bound of crash kernel reservation consistent

[Baoquan He]

On 12/15/21 at 11:01am, Catalin Marinas wrote:
> On Wed, Dec 15, 2021 at 11:42:19AM +0800, Baoquan He wrote:
> > On 12/14/21 at 07:24pm, Catalin Marinas wrote:
> > > On Tue, Dec 14, 2021 at 08:07:58PM +0100, Borislav Petkov wrote:
> > > > On Fri, Dec 10, 2021 at 02:55:25PM +0800, Zhen Lei wrote:
> > > > > From: Chen Zhou <chenzhou10 at huawei.com>
> > > > > 
> > > > > The lower bounds of crash kernel reservation and crash kernel low
> > > > > reservation are different, use the consistent value CRASH_ALIGN.
> > > > 
> > > > A big WHY is missing here to explain why the lower bound of the
> > > > allocation range needs to be 16M and why was 0 wrong?
> > > 
> > > I asked the same here:
> > > 
> > > https://lore.kernel.org/r/20210224143547.GB28965@arm.com
> > > 
> > > IIRC Baoquan said that there is a 1MB reserved for x86 anyway in the
> > > lower part, so that's equivalent in practice to starting from
> > > CRASH_ALIGN.
> > 
> > Yeah, even for i386, there's area reserved by BIOS inside low 1M.
> > Considering the existing alignment CRASH_ALIGN which is 16M, we
> > definitely have no chance to get memory starting from 0. So starting
> > from 16M can skip the useless memblock searching, and make the
> > crashkernel low reservation consisten with crashkernel reservation on
> > allocation code.
> 
> That's the x86 assumption. Is it valid for other architectures once the
> code has been made generic in patch 6? It should be ok for arm64, RAM
> tends to start from higher up but other architectures may start using
> this common code.

Good point. I didn't think of this from generic code side, then let's
keep it as 0.

> 
> If you want to keep the same semantics as before, just leave it as 0.
> It's not that the additional lower bound makes the search slower.

Agree.