Hi Anshuman,
On 7/5/21 1:59 PM, Anshuman Khandual wrote:
On 7/2/21 4:02 PM, Gavin Shan wrote:
There are two addresses selected: random virtual address and physical
address corresponding to kernel symbol @start_kernel. During the PMD
tests in pmd_advanced_tests(), the physical address is aligned down
to the starting address of the huge page, whose size is 512MB on ARM64
when we have 64KB base page size. After that, set_pmd_at() is called
to populate the PMD entry. PG_arch_1, PG_dcache_clean on ARM64, is
set to the page flags. Unforunately, the page, corresponding to the
starting address of the huge page could be owned by buddy. It means
PG_arch_1 can be unconditionally set to page owned by buddy.
Afterwards, the page with PG_arch_1 set is fetched from buddy's free
area list, but fails the checking. It leads to the following warning
on ARM64:
BUG: Bad page state in process memhog pfn:08000
page:0000000015c0a628 refcount:0 mapcount:0 \
mapping:0000000000000000 index:0x1 pfn:0x8000
flags: 0x7ffff8000000800(arch_1|node=0|zone=0|lastcpupid=0xfffff)
raw: 07ffff8000000800 dead000000000100 dead000000000122 0000000000000000
raw: 0000000000000001 0000000000000000 00000000ffffffff 0000000000000000
page dumped because: PAGE_FLAGS_CHECK_AT_PREP flag(s) set
Does this problem happen right after the boot ? OR you ran some tests
and workloads to trigger this ? IIRC never seen this before on arm64.
Does this happen on other archs too ?
The page flag (PG_arch_1) is corrupted during boot on ARM64 where
64KB base page size is selected, but the failing page check happens
when the page is pulled from buddy's free area list by "memhog".
I don't think other platform has same issue.
This fixes the issue by calling flush_dcache_page() after each call
to set_{pud, pmd, pte}_at() because PG_arch_1 isn't needed in any case.
This (arm64 specific solution) might cause some side effects on other
platforms ? The solution here needs to be generic enough. I will take
a look into this patch but probably later this week or next week.
Apart from the overhead of flushing the dcache introduced by flush_dcache_page().
I don't think there is any side-effect. By the way, I'm working on a series
to fix this issue and another issue. I will post the series for review pretty
soon and it's going to fix the following issues:
(1) Current code is organized in relaxed fashion. All information are maintained
in variables in debug_vm_pgtable(). The variables are passed to test functions.
It make the code hard to be maintained in long term. So I will introduce a
dedicated data struct (struct vm_pgtable_debug), as place holder for various
information.
(2) With the data struct, I'm able to allocate page, to be used by set_{pud, pmd, pte}_at()
because the target page is accessed on ARM64. The PG_arch_1 flag is set to
the page and the corresponding iCache is flush if execution permission is given.
There are two issues if the page used by set_{pud, pmd, pte}_at() wasn't allocated
from buddy: (a) the PG_arch_1 flag corruption as this patch tries to fix; (b) kernel
crash because of invalid page fault on accessing the target page. The page isn't
mapped if CONFIG_DEBUG_PAGEALLOC is enabled.
start_kernel
mm_init
mem_init
memblock_free_all
free_low_memory_core_early
__free_memory_core
__free_pages_memory
memblock_free_pages
__free_pages_core
__free_pages_ok
free_pages_prepare
debug_pagealloc_unmap_pages # The page is unmapped here
Thanks,
Gavin
