[RFC v1 02/10] powerpc: book3s64: Fix unmap race with PMD THP migration entry

[Christophe Leroy (CS GROUP)]

Le 25/02/2026 à 12:04, Ritesh Harjani (IBM) a écrit :
> The following race is possible with migration swap entries or
> device-private THP entries. e.g. when move_pages is called on a PMD THP
> page, then there maybe an intermediate state, where PMD entry acts as
> a migration swap entry (pmd_present() is true). Then if an munmap
> happens at the same time, then this VM_BUG_ON() can happen in
> pmdp_huge_get_and_clear_full().
> 
> This patch fixes that.
> 
> Thread A: move_pages() syscall
>    add_folio_for_migration()
>      mmap_read_lock(mm)
>      folio_isolate_lru(folio)
>      mmap_read_unlock(mm)
> 
>    do_move_pages_to_node()
>      migrate_pages()
>        try_to_migrate_one()
>          spin_lock(ptl)
>          set_pmd_migration_entry()
>            pmdp_invalidate()     # PMD: _PAGE_INVALID | _PAGE_PTE | pfn
>            set_pmd_at()          # PMD: migration swap entry (pmd_present=0)
>          spin_unlock(ptl)
>          [page copy phase]       # <--- RACE WINDOW -->
> 
> Thread B: munmap()
>    mmap_write_downgrade(mm)
>    unmap_vmas() -> zap_pmd_range()
>      zap_huge_pmd()
>        __pmd_trans_huge_lock()
>          pmd_is_huge():          # !pmd_present && !pmd_none -> TRUE (swap entry)
>          pmd_lock() -> 		# spin_lock(ptl), waits for Thread A to release ptl
>        pmdp_huge_get_and_clear_full()
>          VM_BUG_ON(!pmd_present(*pmdp))  # HITS!
> 
> [  287.738700][ T1867] ------------[ cut here ]------------
> [  287.743843][ T1867] kernel BUG at arch/powerpc/mm/book3s64/pgtable.c:187!
> cpu 0x0: Vector: 700 (Program Check) at [c00000044037f4f0]
>      pc: c000000000094ca4: pmdp_huge_get_and_clear_full+0x6c/0x23c
>      lr: c000000000645dec: zap_huge_pmd+0xb0/0x868
>      sp: c00000044037f790
>     msr: 800000000282b033
>    current = 0xc0000004032c1a00
>    paca    = 0xc000000004fe0000   irqmask: 0x03   irq_happened: 0x09
>      pid   = 1867, comm = a.out
> kernel BUG at :187!
> Linux version 6.19.0-12136-g14360d4f917c-dirty (powerpc64le-linux-gnu-gcc (Debian 12.2.0-14) 12.2.0, GNU ld (GNU Binutils for Debian) 2.40) #27 SMP PREEMPT Sun Feb 22 10:38:56 IST 2026
> enter ? for help
> [link register   ] c000000000645dec zap_huge_pmd+0xb0/0x868
> [c00000044037f790] c00000044037f7d0 (unreliable)
> [c00000044037f7d0] c000000000645dcc zap_huge_pmd+0x90/0x868
> [c00000044037f840] c0000000005724cc unmap_page_range+0x176c/0x1f40
> [c00000044037fa00] c000000000572ea0 unmap_vmas+0xb0/0x1d8
> [c00000044037fa90] c0000000005af254 unmap_region+0xb4/0x128
> [c00000044037fb50] c0000000005af400 vms_complete_munmap_vmas+0x138/0x310
> [c00000044037fbe0] c0000000005b0f1c do_vmi_align_munmap+0x1ec/0x238
> [c00000044037fd30] c0000000005b3688 __vm_munmap+0x170/0x1f8
> [c00000044037fdf0] c000000000587f74 sys_munmap+0x2c/0x40
> [c00000044037fe10] c000000000032668 system_call_exception+0x128/0x350
> [c00000044037fe50] c00000000000d05c system_call_vectored_common+0x15c/0x2ec
> ---- Exception: 3000 (System Call Vectored) at 0000000010064a2c
> SP (7fff9b1ee9c0) is in userspace
> 0:mon> zh
> 
> Fixes: 75358ea359e7c ("powerpc/mm/book3s64: Fix MADV_DONTNEED and parallel page fault race")
> Reported-by: Pavithra Prakash <pavrampu at linux.vnet.ibm.com>
> Signed-off-by: Ritesh Harjani (IBM) <ritesh.list at gmail.com>

Reviewed-by: Christophe Leroy (CS GROUP) <chleroy at kernel.org>

> ---
>   arch/powerpc/mm/book3s64/pgtable.c | 19 +++++++++++++++++--
>   1 file changed, 17 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/powerpc/mm/book3s64/pgtable.c b/arch/powerpc/mm/book3s64/pgtable.c
> index 4b09c04654a8..359092001670 100644
> --- a/arch/powerpc/mm/book3s64/pgtable.c
> +++ b/arch/powerpc/mm/book3s64/pgtable.c
> @@ -210,8 +210,23 @@ pmd_t pmdp_huge_get_and_clear_full(struct vm_area_struct *vma,
>   {
>   	pmd_t pmd;
>   	VM_BUG_ON(addr & ~HPAGE_PMD_MASK);
> -	VM_BUG_ON((pmd_present(*pmdp) && !pmd_trans_huge(*pmdp)) ||
> -		   !pmd_present(*pmdp));
> +	VM_BUG_ON((pmd_present(*pmdp) && !pmd_trans_huge(*pmdp)));
> +
> +	if (!pmd_present(*pmdp)) {
> +		/*
> +		 * Non-present PMDs can be migration entries or device-private
> +		 * THP entries. Since these are non-present, so there is no TLB
> +		 * backing. This happens when the address space is being
> +		 * unmapped zap_huge_pmd(), and we encounter non-present pmds.
> +		 * So it is safe to just clear the PMDs here. zap_huge_pmd(),
> +		 * will take care of withdraw of the deposited table.
> +		 */
> +		pmd = pmdp_get(pmdp);
> +		pmd_clear(pmdp);
> +		page_table_check_pmd_clear(vma->vm_mm, addr, pmd);
> +		return pmd;
> +	}
> +
>   	pmd = pmdp_huge_get_and_clear(vma->vm_mm, addr, pmdp);
>   	/*
>   	 * if it not a fullmm flush, then we can possibly end up converting
> --
> 2.53.0
>