On Fri, May 07, 2021 at 03:17:08PM +0800, Kefeng Wang wrote:

On 2021/5/6 20:47, Kefeng Wang wrote:

quoted

quoted

quoted

quoted

quoted

no, the CONFIG_ARM_LPAE is not set, and yes with same panic at
move_freepages at

start_pfn/end_pfn [de600, de7ff], [de600000, de7ff000]
:  pfn =de600, page
=ef3cc000, page-flags = ffffffff,  pfn2phy = de600000

quoted

quoted

__free_memory_core, range: 0xb0200000 -
0xc0000000, pfn: b0200 - b0200
__free_memory_core, range: 0xcc000000 -
0xdca00000, pfn: cc000 - b0200
__free_memory_core, range: 0xde700000 -
0xdea00000, pfn: de700 - b0200

Hmm, [de600, de7ff] is not added to the free lists which is
correct. But
then it's unclear how the page for de600 gets to move_freepages()...

Can't say I have any bright ideas to try here...

Are we missing some checks (e.g., PageReserved()) that
pfn_valid_within()
would have "caught" before?

Unless I'm missing something the crash happens in __rmqueue_fallback():

do_steal:
    page = get_page_from_free_area(area, fallback_mt);

    steal_suitable_fallback(zone, page, alloc_flags, start_migratetype,
                                can_steal);
        -> move_freepages()
            -> BUG()

So a page from free area should be sane as the freed range was never
added
it to the free lists.

Sorry for the late response due to the vacation.

The pfn in range [de600, de7ff] won't be added into the free lists via
__free_memory_core(), but the pfn could be added into freelists via
free_highmem_page()

I add some debug[1] in add_to_free_list(), we could see the calltrace

free_highpages, range_pfn [b0200, c0000], range_addr [b0200000, c0000000]
free_highpages, range_pfn [cc000, dca00], range_addr [cc000000, dca00000]
free_highpages, range_pfn [de700, dea00], range_addr [de700000, dea00000]
add_to_free_list, ===> pfn = de700
------------[ cut here ]------------
WARNING: CPU: 0 PID: 0 at mm/page_alloc.c:900 add_to_free_list+0x8c/0xec
pfn = de700
Modules linked in:
CPU: 0 PID: 0 Comm: swapper Not tainted 5.10.0+ #48
Hardware name: Hisilicon A9
[<c010a600>] (show_stack) from [<c04b21c4>] (dump_stack+0x9c/0xc0)
[<c04b21c4>] (dump_stack) from [<c011c708>] (__warn+0xc0/0xec)
[<c011c708>] (__warn) from [<c011c7a8>] (warn_slowpath_fmt+0x74/0xa4)
[<c011c7a8>] (warn_slowpath_fmt) from [<c023721c>]
(add_to_free_list+0x8c/0xec)
[<c023721c>] (add_to_free_list) from [<c0237e00>]
(free_pcppages_bulk+0x200/0x278)
[<c0237e00>] (free_pcppages_bulk) from [<c0238d14>]
(free_unref_page+0x58/0x68)
[<c0238d14>] (free_unref_page) from [<c023bb54>]
(free_highmem_page+0xc/0x50)
[<c023bb54>] (free_highmem_page) from [<c070620c>] (mem_init+0x21c/0x254)
[<c070620c>] (mem_init) from [<c0700b38>] (start_kernel+0x258/0x5c0)
[<c0700b38>] (start_kernel) from [<00000000>] (0x0)

so any idea?

If pfn = 0xde700, due to the pageblock_nr_pages = 0x200, then the
start_pfn,end_pfn passed to move_freepages() will be [de600, de7ff],
but the range of [de600,de700] without ‘struct page' will lead to
this panic when pfn_valid_within not enabled if no HOLES_IN_ZONE,
and the same issue will occurred in isolate_freepages_block(), maybe

I think your analysis is correct except one minor detail. With the #ifdef
fix I've proposed earlieri [1] the memmap for [0xde600, 0xde700] should not
be freed so there should be a struct page. Did you check what parts of the
memmap are actually freed with this patch applied?
Would you get a panic if you add

	dump_page(pfn_to_page(0xde600), "");

say, in the end of memblock_free_all()?

there are some scene, so I select HOLES_IN_ZONE in ARCH_HISI(ARM) to solve
this issue in our 5.10, should we select HOLES_IN_ZONE in all ARM or only in
ARCH_HISI, any better solution?  Thanks.

I don't think that HOLES_IN_ZONE is the right solution. I believe that we
must keep the memory map aligned on pageblock boundaries. That's surely not the
case for SPARSEMEM as of now, and if my fix is not enough we need to find
where it went wrong.

Besides, I'd say that if it is possible to update your firmware to make the
memory layout reported to the kernel less, hmm, esoteric, you would hit
less corner cases.

[1] https://lore.kernel.org/lkml/YIpY8TXCSc7Lfa2Z@kernel.org (local)

-- 
Sincerely yours,
Mike.