[permalink] [raw]

Subject: Re: [PATCH v2 4/6] mm/mm_init.c: remove meaningless calculation of zone->managed_pages in free_area_init_core()

On 03/28/24 at 11:53am, Mike Rapoport wrote:
> On Thu, Mar 28, 2024 at 04:32:38PM +0800, Baoquan He wrote:
> > On 03/25/24 at 10:56pm, Baoquan He wrote:
> > >
> > > /*
> > > - * Set an approximate value for lowmem here, it will be adjusted
> > > - * when the bootmem allocator frees pages into the buddy system.
> > > - * And all highmem pages will be managed by the buddy system.
> > > + * Initialize zone->managed_pages as 0 , it will be reset
> > > + * when memblock allocator frees pages into buddy system.
> > > */
> > > - zone_init_internals(zone, j, nid, freesize);
> > > + zone_init_internals(zone, j, nid, 0);
> >
> > Here, we should initialize zone->managed_pages as zone->present_pages
> > because later page_group_by_mobility_disabled need be set according to
> > zone->managed_pages. Otherwise page_group_by_mobility_disabled will be
> > set to 1 always. I will sent out v3.
>
> With zone->managed_pages set to zone->present_pages we won't account for
> the reserved memory for initialization of page_group_by_mobility_disabled.

The old zone->managed_pages didn't account for the reserved pages
either. It's calculated by (zone->present_pages - memmap_pages). memmap
pages only is only a very small portion, e.g on x86_64, 4K page size,
assuming size of struct page is 64, then it's 1/64 of system memory.
On arm64, 64K page size, it's 1/1024 of system memory.

And about the setting of page_group_by_mobility_disabled, the compared
value pageblock_nr_pages * MIGRATE_TYPES which is very small. On x86_64,
it's 4M*6=24M; on arm64 with 64K size and 128M*6=768M which should be
the biggest among ARCH-es.

if (vm_total_pages < (pageblock_nr_pages * MIGRATE_TYPES))
page_group_by_mobility_disabled = 1;
else
page_group_by_mobility_disabled = 0;

So page_group_by_mobility_disabled could be set to 1 only on system with
very little memory which is very rarely seen. And setting
zone->managed_pages as zone->present_pages is very close to its old
value: (zone->present_pages - memmap_pages). Here we don't need be very
accurate, just a rough value.

>
> As watermarks are still not initialized at the time build_all_zonelists()
> is called, we may use nr_all_pages - nr_kernel_pages instead of
> nr_free_zone_pages(), IMO.

nr_all_pages should be fine if we take this way. nr_kernel_pages is a
misleading name, it's all low memory pages excluding kernel reserved
apges. nr_all_pages is all memory pages including highmema and exluding
kernel reserved pages.

Both is fine to me. The first one is easier, simply setting
zone->managed_pages as zone->present_pages. The 2nd way is a little more
accurate.

>
> > From a17b0921b4bd00596330f61ee9ea4b82386a9fed Mon Sep 17 00:00:00 2001
> > From: Baoquan He <[email protected]>
> > Date: Thu, 28 Mar 2024 16:20:15 +0800
> > Subject: [PATCH] mm/mm_init.c: set zone's ->managed_pages as ->present_pages
> > for now
> > Content-type: text/plain
> >
> > Because page_group_by_mobility_disabled need be set according to zone's
> > managed_pages later.
> >
> > Signed-off-by: Baoquan He <[email protected]>
> > ---
> > mm/mm_init.c | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/mm/mm_init.c b/mm/mm_init.c
> > index cc24e7958c0c..dd875f943cbb 100644
> > --- a/mm/mm_init.c
> > +++ b/mm/mm_init.c
> > @@ -1561,7 +1561,7 @@ static void __init free_area_init_core(struct pglist_data *pgdat)
> > * Initialize zone->managed_pages as 0 , it will be reset
> > * when memblock allocator frees pages into buddy system.
> > */
> > - zone_init_internals(zone, j, nid, 0);
> > + zone_init_internals(zone, j, nid, zone->present_pages);
> >
> > if (!size)
> > continue;
> > --
> > 2.41.0
> >
> >
> > >
> > > if (!size)
> > > continue;
> > > @@ -1915,6 +1878,7 @@ void __init free_area_init(unsigned long *max_zone_pfn)
> > > check_for_memory(pgdat);
> > > }
> > >
> > > + calc_nr_kernel_pages();
> > > memmap_init();
> > >
> > > /* disable hash distribution for systems with a single node */
> > > --
> > > 2.41.0
> > >
> >
>
> --
> Sincerely yours,
> Mike.
>