On 2016/12/16 2:31, Andrea Reale wrote:

Hi Xishi Qiu,

thanks for your comments. 

The short anwser to your question is the following.  As you hinted,
it is related to the way pfn_valid() is implemented in arm64 when
CONFIG_HAVE_ARCH_PFN_VALID is true (default), i.e., just a check for
the NOMAP flags on the corresponding memblocks.

Since arch_add_memory->__add_pages() expects pfn_valid() to return false
when it is first called, we mark corresponding memory blocks with NOMAP;
however, arch_add_memory->__add_pages()->__add_section()->__add_zone()
expects pfn_valid() to return true when, at the end of its body,
it cycles through pages to call SetPageReserved(). Since blocks are
marked with NOMAP, pages will not be reserved there, henceforth we
need to reserve them after we clear the NOMAP flag inside the body of
arch_add_memory. Having pages reserved at the end of arch_add_memory
is a preconditions for the upcoming onlining of memory blocks (see
memory_block_change_state()).

quoted

It's because that in memmap_init_zone() -> early_pfn_valid(), the new page is still
invalid, so we need to init it after memblock_clear_nomap()

So why not use __init_single_page() and set_pageblock_migratetype()?

About your comment on memmap_init_zone()->early_pfn_valid(), I think
that that particular check is never executed in the context of memory
hotplug; in fact, just before the call to early_pfn_valid(), the code
will jump to the `not_early` label, because the context == MEMMAP_HOPTLUG.

Hi Andrea,

Thanks for your answer, I notice it now. I thought SetPageReserved() was
done in memmap_init_zone(), but the new kernel change this logic now, so
sorry for the noise.

Thanks,
Xishi Qiu