Re: [PATCH] memblock: update numa node of memblk reserved type

[Kefeng Wang]

On 2023/5/24 12:59, Anshuman Khandual wrote:
> On 5/23/23 17:27, Kefeng Wang wrote:
> > The numa node of memblk reserved type is wrong, it could update
> > according to the numa node information from memblk memory type,
> > let's fix it.
>
> Indeed it's wrong at present and can be verified from sysfs file
> (/sys/kernel/debug/memblock/reserved) accessed in user space.

Yes, both memblock_dump() and sysfs show wrong value.
> > Signed-off-by: Kefeng Wang <wangkefeng.wang@xxxxxxxxxx>
> > ---
> >   mm/memblock.c | 25 +++++++++++++++++++++++++
> >   1 file changed, 25 insertions(+)
> >
> > diff --git a/mm/memblock.c b/mm/memblock.c
> > index a50447d970ef..45a0781cda31 100644
> > --- a/mm/memblock.c
> > +++ b/mm/memblock.c
> > @@ -1922,6 +1922,28 @@ phys_addr_t __init_memblock memblock_get_current_limit(void)
> >   	return memblock.current_limit;
> >   }
> >   
> > +static void __init_memblock memblock_reserved_update_node(void)
> > +{
> > +	struct memblock_region *rgn;
> > +	phys_addr_t base, end, size;
> > +	int ret;
> > +
> > +	if (!IS_ENABLED(CONFIG_NUMA))
> > +		return;
> > +
> > +	for_each_mem_region(rgn) {
> > +		base = rgn->base;
> > +		size = rgn->size;
> > +		end = base + size - 1;
> > +
> > +		ret = memblock_set_node(base, size, &memblock.reserved,
> > +					memblock_get_region_node(rgn));
> > +		if (ret)
> > +			pr_err("memblock: Failed to update reserved [%pa-%pa] node",
> > +			       &base, &end);
> > +	}
> > +}
> > +
> >   static void __init_memblock memblock_dump(struct memblock_type *type)
> >   {
> >   	phys_addr_t base, end, size;
> > @@ -1955,6 +1977,7 @@ static void __init_memblock __memblock_dump_all(void)
> >   		&memblock.memory.total_size,
> >   		&memblock.reserved.total_size);
> >   
> > +	memblock_reserved_update_node();
>
> __memblock_dump_all() gets called only when memblock_debug is enabled.
> This helper should be called directly inside memblock_dump_all() right
> at the beginning, regardless of memblock_debug.

This is my first though, but I found there are still many memblock_alloc 
and memblock_reserve after memblock_dump_all(), so I update it twice,

1) __memblock_dump_all()
2) memblock_debug_show()

and without the above two interface, no one care about the reserved node
info, so I put memblock_reserved_update_node into __memblock_dump_all().


> >   #ifdef CONFIG_HAVE_MEMBLOCK_PHYS_MAP
> > @@ -2196,6 +2219,8 @@ static int memblock_debug_show(struct seq_file *m, void *private)
> >   	unsigned int count = ARRAY_SIZE(flagname);
> >   	phys_addr_t end;
> >   
> > +	memblock_reserved_update_node();
> > +

> This is redundant, should be dropped. Reserved memblock ranges need not
> be scanned, each time the sysfs file is accessed from user space.

Yes, it's better to move it into memblock_init_debugfs(),
which only called once.