On Mon, 13 Jun 2022 at 13:05, Mike Rapoport <rppt@xxxxxxxxxx> wrote:
>
> On Mon, Jun 13, 2022 at 04:21:47PM +0800, Wupeng Ma wrote:
> > From: Ma Wupeng <mawupeng1@xxxxxxxxxx>
> >
> > If system have some mirrored memory and mirrored feature is not specified
> > in boot parameter, the basic mirrored feature will be enabled and this will
> > lead to the following situations:
> >
> > - memblock memory allocation prefers mirrored region. This may have some
> > unexpected influence on numa affinity.
> >
> > - contiguous memory will be split into several parts if parts of them
> > is mirrored memory via memblock_mark_mirror().
> >
> > To fix this, variable mirrored_kernelcore will be checked in
> > memblock_mark_mirror(). Mark mirrored memory with flag MEMBLOCK_MIRROR iff
> > kernelcore=mirror is added in the kernel parameters.
> >
> > Signed-off-by: Ma Wupeng <mawupeng1@xxxxxxxxxx>
> > ---
> > mm/internal.h | 2 ++
> > mm/memblock.c | 3 +++
> > mm/page_alloc.c | 2 +-
> > 3 files changed, 6 insertions(+), 1 deletion(-)
> >
> > diff --git a/mm/internal.h b/mm/internal.h
> > index c0f8fbe0445b..ddd2d6a46f1b 100644
> > --- a/mm/internal.h
> > +++ b/mm/internal.h
> > @@ -861,4 +861,6 @@ struct folio *try_grab_folio(struct page *page, int refs, unsigned int flags);
> >
> > DECLARE_PER_CPU(struct per_cpu_nodestat, boot_nodestats);
> >
> > +extern bool mirrored_kernelcore;
> > +
> > #endif /* __MM_INTERNAL_H */
> > diff --git a/mm/memblock.c b/mm/memblock.c
> > index b1d2a0009733..a9f18b988b7f 100644
> > --- a/mm/memblock.c
> > +++ b/mm/memblock.c
> > @@ -924,6 +924,9 @@ int __init_memblock memblock_clear_hotplug(phys_addr_t base, phys_addr_t size)
> > */
> > int __init_memblock memblock_mark_mirror(phys_addr_t base, phys_addr_t size)
> > {
> > + if (!mirrored_kernelcore)
> > + return 0;
> > +
>
> Hmm, this changes the way x86 uses mirrored memory.
> This change makes sense for x86 as well, but we should get an Ack from x86 folks.
>
Also, on second thought, I don't think marking as mirror is what
should be affected by the policy. Instead, choose_memblock_flags()
should take this into account, in a way that we could refine later if
needed.
