Re: [PATCH] arm64: dax: add devmap check for pmd_trans_huge

[Alistair Popple <apopple@nvidia.com>]

On Tue, Apr 08, 2025 at 04:59:14PM +0800, Wupeng Ma wrote:
> During our test in ext4 dax linux-v5.10 on arm64. A BUG_ON is trigger in
> follow_invalidate_pte as follow since this pmd is seem as pmd_trans_huge.
> However this page is really a dax-pmds rather than a pmd trans huge.
> 
> Call trace is shown as follow:
> 
>   ------------[ cut here ]------------
>   kernel BUG at mm/memory.c:5185!
>   CPU: 0 PID: 150 Comm: kworker/u8:10 Not tainted 5.10.0-01678-g1e62aad66bbc-dirty #36

This is an old kernel, and I couldn't correlate the line number of the BUG_ON()
probably because you have patches applied. But I assume this is the VM_BUG_ON()
in follow_invalidate_pte()? Does this issue reproduce on more recent kernel
versions (eg. v6.13)? Or some other upstream kernel version?

>   pc : follow_invalidate_pte+0xdc/0x5e0
>   lr : follow_invalidate_pte+0xc4/0x5e0
>   sp : ffffa00012997110
>   Call trace:
>    follow_invalidate_pte+0xdc/0x5e0
>    dax_entry_mkclean+0x250/0x870
>    dax_writeback_one+0xac/0x380
>    dax_writeback_mapping_range+0x22c/0x704
>    ext4_dax_writepages+0x234/0x6e4
>    do_writepages+0xc8/0x1c0
>    __writeback_single_inode+0xb8/0x560
>    writeback_sb_inodes+0x344/0x7a0
>    wb_writeback+0x1f8/0x6b0
>    wb_do_writeback+0x194/0x3cc
>    wb_workfn+0x14c/0x590
>    process_one_work+0x470/0xa30
>    worker_thread+0xac/0x510
>    kthread+0x1e0/0x220
>    ret_from_fork+0x10/0x18
>   ---[ end trace 0f479050bd4b1818 ]---
>   Kernel panic - not syncing: Oops - BUG: Fatal exception
>   ---[ end Kernel panic - not syncing: Oops - BUG: Fatal exception ]---
> 
> Commit 5c7fb56e5e3f ("mm, dax: dax-pmd vs thp-pmd vs hugetlbfs-pmd") and
> commit 36b78402d97a ("powerpc/hash64/devmap: Use H_PAGE_THP_HUGE when
> setting up huge devmap PTE entries") already check pmd_devmap during
> checking pmd_trans_huge. Since pmd_devmap() is used to distinguish dax-pmds,
> add the same check for arm64 to fix this problem.

That seems correct to me. In practice most callers of pmd_trans_huge() that can
see a dax-pmd already check for it explicitly with vma_is_dax(), but there are a
few cases that don't.

> Add PTE_DEVMAP in pte_modify as commit 4628a64591e6 ("mm: Preserve
> _PAGE_DEVMAP across mprotect() calls") does to avoid the same issue in
> mprotect.
> 
> Fixes: 73b20c84d42d ("arm64: mm: implement pte_devmap support")
> Signed-off-by: Wupeng Ma <mawupeng1@huawei.com>
> ---
>  arch/arm64/include/asm/pgtable.h | 5 +++--
>  1 file changed, 3 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
> index d3b538be1500b..b9a618127c01b 100644
> --- a/arch/arm64/include/asm/pgtable.h
> +++ b/arch/arm64/include/asm/pgtable.h
> @@ -740,7 +740,7 @@ static inline int pmd_trans_huge(pmd_t pmd)
>  	 * as a table, so force the valid bit for the comparison.
>  	 */
>  	return pmd_val(pmd) && pmd_present(pmd) &&
> -	       !pmd_table(__pmd(pmd_val(pmd) | PTE_VALID));
> +	       !pmd_table(__pmd(pmd_val(pmd) | PTE_VALID)) && !pmd_devmap(pmd);
>  }
>  #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
>  
> @@ -1186,7 +1186,8 @@ static inline pte_t pte_modify(pte_t pte, pgprot_t newprot)
>  	 */
>  	const pteval_t mask = PTE_USER | PTE_PXN | PTE_UXN | PTE_RDONLY |
>  			      PTE_PRESENT_INVALID | PTE_VALID | PTE_WRITE |
> -			      PTE_GP | PTE_ATTRINDX_MASK | PTE_PO_IDX_MASK;
> +			      PTE_GP | PTE_ATTRINDX_MASK | PTE_PO_IDX_MASK |
> +			      PTE_DEVMAP;
>  
>  	/* preserve the hardware dirty information */
>  	if (pte_hw_dirty(pte))
> -- 
> 2.43.0
> 
>