Re: [PATCH 1/1] mm: Always release pages to the buddy allocator in memblock_free_late().

From: Aaron Thompson
Date: Wed Jan 04 2023 - 23:41:59 EST


Hi Mike,

On 2023-01-04 11:34, Mike Rapoport wrote:
Hi,

On Wed, Jan 04, 2023 at 07:43:36AM +0000, Aaron Thompson wrote:
If CONFIG_DEFERRED_STRUCT_PAGE_INIT is enabled, memblock_free_pages()
only releases pages to the buddy allocator if they are not in the
deferred range. This is correct for free pages (as defined by
for_each_free_mem_pfn_range_in_zone()) because free pages in the
deferred range will be initialized and released as part of the deferred
init process. memblock_free_pages() is called by memblock_free_late(),
which is used to free reserved ranges after memblock_free_all() has
run. memblock_free_all() initializes all pages in reserved ranges, and

To be precise, memblock_free_all() frees pages, or releases them to the
pages allocator, rather than initializes.

As you mentioned in the comment below, whether memblock_free_all() does any
initializing depends on the particular deferred init situation.
memblock_free_all() does ultimately call init_reserved_page() for every reserved
page (via reserve_bootmem_region()), but that only actually initializes the page
if it's in the deferred range. In either case, all I was trying to say here is
that we can be certain that all reserved pages have been initialized after
memblock_free_all() has run, so I'll rephrase that.

accordingly, those pages are not touched by the deferred init
process. This means that currently, if the pages that
memblock_free_late() intends to release are in the deferred range, they
will never be released to the buddy allocator. They will forever be
reserved.

In addition, memblock_free_pages() calls kmsan_memblock_free_pages(),
which is also correct for free pages but is not correct for reserved
pages. KMSAN metadata for reserved pages is initialized by
kmsan_init_shadow(), which runs shortly before memblock_free_all().

For both of these reasons, memblock_free_pages() should only be called
for free pages, and memblock_free_late() should call __free_pages_core()
directly instead.

Overall looks fine to me and I couldn't spot potential issues.

I'd appreciate if you add a paragraph about the actual issue with EFI boot
you described in the cover letter to the commit message.

Sure, will do.

Fixes: 3a80a7fa7989 ("mm: meminit: initialise a subset of struct pages if CONFIG_DEFERRED_STRUCT_PAGE_INIT is set")
Signed-off-by: Aaron Thompson <dev@xxxxxxxxxx>
---
mm/memblock.c | 2 +-
tools/testing/memblock/internal.h | 4 ++++
2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/mm/memblock.c b/mm/memblock.c
index 511d4783dcf1..56a5b6086c50 100644
--- a/mm/memblock.c
+++ b/mm/memblock.c
@@ -1640,7 +1640,7 @@ void __init memblock_free_late(phys_addr_t base, phys_addr_t size)
end = PFN_DOWN(base + size);

for (; cursor < end; cursor++) {
- memblock_free_pages(pfn_to_page(cursor), cursor, 0);
+ __free_pages_core(pfn_to_page(cursor), 0);

Please add a comment that explains why it is safe to call
__free_pages_core() here.
Something like

/*
* Reserved pages are always initialized by the end of
* memblock_free_all() either during memmap_init() or, with deferred
* initialization if struct page in reserve_bootmem_region()
*/

Will do. Thanks for the review.

totalram_pages_inc();
}
}
diff --git a/tools/testing/memblock/internal.h b/tools/testing/memblock/internal.h
index fdb7f5db7308..85973e55489e 100644
--- a/tools/testing/memblock/internal.h
+++ b/tools/testing/memblock/internal.h
@@ -15,6 +15,10 @@ bool mirrored_kernelcore = false;

struct page {};

+void __free_pages_core(struct page *page, unsigned int order)
+{
+}
+
void memblock_free_pages(struct page *page, unsigned long pfn,
unsigned int order)
{
--
2.30.2


Thanks,
-- Aaron