Re: [PATCH] mm/sparse: Consistently do not zero memmap

[Vincent Whitchurch <vincent.whitchurch@axis.com>]

On Wed, Oct 30, 2019 at 02:29:58PM +0100, Michal Hocko wrote:
> On Wed 30-10-19 14:11:22, Vincent Whitchurch wrote:
> > (I noticed this because on my ARM64 platform, with 1 GiB of memory the
> >  first [and only] section is allocated from the zeroing path while with
> >  2 GiB of memory the first 1 GiB section is allocated from the
> >  non-zeroing path.)
> 
> Do I get it right that sparse_buffer_init couldn't allocate memmap for
> the full node for some reason and so sparse_init_nid would have to
> allocate one for each memory section?

Not quite.  The sparsemap_buf is successfully allocated with the correct
size in sparse_buffer_init(), but sparse_buffer_alloc() fails to
allocate the same size from it.

The reason it fails is that sparse_buffer_alloc() for some reason wants
to return a pointer which is aligned to the allocation size.  But the
sparsemap_buf was only allocated with PAGE_SIZE alignment so there's not
enough space to align it.

I don't understand the reason for this alignment requirement since the
fallback path also allocates with PAGE_SIZE alignment.  I'm guessing the
alignment is for the VMEMAP code which also uses sparse_buffer_alloc()?