[PATCH v3 5/6] mm: Add mirror flag back on initrd memory

[Ard Biesheuvel]

On Tue, 7 Jun 2022 at 14:22, David Hildenbrand <david at redhat.com> wrote:
>
> On 07.06.22 11:38, Wupeng Ma wrote:
> > From: Ma Wupeng <mawupeng1 at huawei.com>
> >
> > Initrd memory will be removed and then added in arm64_memblock_init() and this
> > will cause it to lose all of its memblock flags. The lost of MEMBLOCK_MIRROR
> > flag will lead to error log printed by find_zone_movable_pfns_for_nodes if
> > the lower 4G range has some non-mirrored memory.
> >
> > In order to solve this problem, the lost MEMBLOCK_MIRROR flag will be
> > reinstalled if the origin memblock has this flag.
> >
> > Signed-off-by: Ma Wupeng <mawupeng1 at huawei.com>
> > ---
> >  arch/arm64/mm/init.c     |  9 +++++++++
> >  include/linux/memblock.h |  1 +
> >  mm/memblock.c            | 20 ++++++++++++++++++++
> >  3 files changed, 30 insertions(+)
> >
> > diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c
> > index 339ee84e5a61..11641f924d08 100644
> > --- a/arch/arm64/mm/init.c
> > +++ b/arch/arm64/mm/init.c
> > @@ -350,9 +350,18 @@ void __init arm64_memblock_init(void)
> >                       "initrd not fully accessible via the linear mapping -- please check your bootloader ...\n")) {
> >                       phys_initrd_size = 0;
> >               } else {
> > +                     int flags, ret;
> > +
> > +                     ret = memblock_get_flags(base, &flags);
> > +                     if (ret)
> > +                             flags = 0;
> > +
> >                       memblock_remove(base, size); /* clear MEMBLOCK_ flags */
> >                       memblock_add(base, size);
> >                       memblock_reserve(base, size);
>
> Can you explain why we're removing+re-adding here exactly? Is it just to
> clear flags as the comment indicates?
>

This should only happen if the placement of the initrd conflicts with
a mem= command line parameter or it is not covered by memblock for
some other reason.

IOW, this should never happen, and if re-memblock_add'ing this memory
unconditionally is causing problems, we should fix that instead of
working around it.

> If it's really just about clearing flags, I wonder if we rather want to
> have an interface that does exactly that, and hides the way this is
> actually implemented (obtain flags, remove, re-add ...), internally.
>
> But most probably there is more magic in the code and clearing flags
> isn't all it ends up doing.
>

I don't remember exactly why we needed to clear the flags, but I think
it had to do with some corner case we hit when the initrd was
partially covered.