Re: [PATCH v2 0/4] arm64: drop pfn_valid_within() and simplify pfn_valid()

[Mike Rapoport <rppt@kernel.org>]

On Thu, Apr 22, 2021 at 11:28:24PM +0800, Kefeng Wang wrote:
> 
> On 2021/4/22 15:29, Mike Rapoport wrote:
> > On Thu, Apr 22, 2021 at 03:00:20PM +0800, Kefeng Wang wrote:
> > > On 2021/4/21 14:51, Mike Rapoport wrote:
> > > > From: Mike Rapoport <rppt@linux.ibm.com>
> > > > 
> > > > Hi,
> > > > 
> > > > These patches aim to remove CONFIG_HOLES_IN_ZONE and essentially hardwire
> > > > pfn_valid_within() to 1.
> > > > 
> > > > The idea is to mark NOMAP pages as reserved in the memory map and restore
> > > > the intended semantics of pfn_valid() to designate availability of struct
> > > > page for a pfn.
> > > > 
> > > > With this the core mm will be able to cope with the fact that it cannot use
> > > > NOMAP pages and the holes created by NOMAP ranges within MAX_ORDER blocks
> > > > will be treated correctly even without the need for pfn_valid_within.
> > > > 
> > > > The patches are only boot tested on qemu-system-aarch64 so I'd really
> > > > appreciate memory stress tests on real hardware.
> > > > 
> > > > If this actually works we'll be one step closer to drop custom pfn_valid()
> > > > on arm64 altogether.
> > > Hi Mike，I have a question, without HOLES_IN_ZONE, the pfn_valid_within() in
> > > move_freepages_block()->move_freepages()
> > > will be optimized, if there are holes in zone, the 'struce page'(memory map)
> > > for pfn range of hole will be free by
> > > free_memmap(), and then the page traverse in the zone(with holes) from
> > > move_freepages() will meet the wrong page，
> > > then it could panic at PageLRU(page) test, check link[1],
> > First, HOLES_IN_ZONE name us hugely misleading, this configuration option
> > has nothing to to with memory holes, but rather it is there to deal with
> > holes or undefined struct pages in the memory map, when these holes can be
> > inside a MAX_ORDER_NR_PAGES region.
> > 
> > In general pfn walkers use pfn_valid() and pfn_valid_within() to avoid
> > accessing *missing* struct pages, like those that are freed at
> > free_memmap(). But on arm64 these tests also filter out the nomap entries
> > because their struct pages are not initialized.
> > 
> > The panic you refer to happened because there was an uninitialized struct
> > page in the middle of MAX_ORDER_NR_PAGES region because it corresponded to
> > nomap memory.
> > 
> > With these changes I make sure that such pages will be properly initialized
> > as PageReserved and the pfn walkers will be able to rely on the memory map.
> > 
> > Note also, that free_memmap() aligns the parts being freed on MAX_ORDER
> > boundaries, so there will be no missing parts in the memory map within a
> > MAX_ORDER_NR_PAGES region.
> 
> Ok, thanks, we met a same panic like the link on arm32(without
> HOLES_IN_ZONE),
> 
> the scheme for arm64 could be suit for arm32, right?

In general yes. You just need to make sure that usage of pfn_valid() in
arch/arm does not presume that it tests something beyond availability of
struct page for a pfn.
 
> I will try the patchset with some changes on arm32 and give some
> feedback.
> 
> Again, the stupid question, where will mark the region of memblock with
> MEMBLOCK_NOMAP flag ?
 
Not sure I understand the question. The memory regions with "nomap"
property in the device tree will be marked MEMBLOCK_NOMAP.
 
> > > "The idea is to mark NOMAP pages as reserved in the memory map", I see the
> > > patch2 check memblock_is_nomap() in memory region
> > > of memblock, but it seems that memblock_mark_nomap() is not called(maybe I
> > > missed), then memmap_init_reserved_pages() won't
> > > work, so should the HOLES_IN_ZONE still be needed for generic mm code?
> > > 
> > > [1] https://lore.kernel.org/linux-arm-kernel/541193a6-2bce-f042-5bb2-88913d5f1047@arm.com/
> > > 

-- 
Sincerely yours,
Mike.