On Thu, 21 Apr 2022 at 00:27, Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> wrote: > > On Tue, 19 Apr 2022 20:29:27 +0200 Ard Biesheuvel <ardb@xxxxxxxxxx> wrote: > > > > > --- a/mm/page_alloc.c > > > > +++ b/mm/page_alloc.c > > > > @@ -7870,7 +7870,7 @@ static void __init find_zone_movable_pfns_for_nodes(void) > > > > > > > > usable_startpfn = memblock_region_memory_base_pfn(r); > > > > > > > > - if (usable_startpfn < 0x100000) { > > > > + if (usable_startpfn < PHYS_PFN(SZ_4G)) { > > > > mem_below_4gb_not_mirrored = true; > > > > continue; > > > > } > > > > > > Regardless PFN value should never be encoded directly. > > > > > > Reviewed-by: Anshuman Khandual <anshuman.khandual@xxxxxxx> > > > > Acked-by: Ard Biesheuvel <ardb@xxxxxxxxxx> > > > > Andrew, can you please take this one through the -mm tree? The rest of > > the series needs a bit more work, but is an obvious fix and there is > > no point in holding it up. > > Sure. > > I'm not seeing any description of the runtime effects of this > shortcoming. I tentatively queued the fix for 5.18, without a > cc:stable for backporting. But that might not be the best decision? > As far as I can tell, mirrored memory is only used on x86 today, where pages are always 4k. However, the whole notion of memory below 4 GB being special is a x86-ism, and so this logic does not appear to extrapolate to other architectures anyway, and probably needs more work. So definitely not a backportable fix, but just an incremental improvement, so either 5.18 or 5.19 should be fine afaict (and no cc:stable)