V1 -> V2: - Because `pmd_val` variable broke ppc builds due to its name, renamed it to `_pmd`. see [1]. [1] https://lore.kernel.org/stable/aS7lPZPYuChOTdXU@hyeyoo
- Added David Hildenbrand's Acked-by [2], thanks a lot! [2] https://lore.kernel.org/linux-mm/ac8d7137-3819-4a75-9dd3-fb3d2259ebe4@kernel...
# TL;DR
previous discussion: https://lore.kernel.org/linux-mm/20250921232709.1608699-1-harry.yoo@oracle.c...
A "bad pmd" error occurs due to race condition between change_prot_numa() and THP migration. The mainline kernel does not have this bug as commit 670ddd8cdc fixes the race condition. 6.1.y, 5.15.y, 5.10.y, 5.4.y are affected by this bug.
Fixing this in -stable kernels is tricky because pte_map_offset_lock() has different semantics in pre-6.5 and post-6.5 kernels. I am trying to backport the same mechanism we have in the mainline kernel. Since the code looks bit different due to different semantics of pte_map_offset_lock(), it'd be best to get this reviewed by MM folks.
# Testing
I verified that the bug described below is not reproduced anymore (on a downstream kernel) after applying this patch series. It used to trigger in few days of intensive numa balancing testing, but it survived 2 weeks with this applied.
# Bug Description
It was reported that a bad pmd is seen when automatic NUMA balancing is marking page table entries as prot_numa:
[2437548.196018] mm/pgtable-generic.c:50: bad pmd 00000000af22fc02(dffffffe71fbfe02) [2437548.235022] Call Trace: [2437548.238234] <TASK> [2437548.241060] dump_stack_lvl+0x46/0x61 [2437548.245689] panic+0x106/0x2e5 [2437548.249497] pmd_clear_bad+0x3c/0x3c [2437548.253967] change_pmd_range.isra.0+0x34d/0x3a7 [2437548.259537] change_p4d_range+0x156/0x20e [2437548.264392] change_protection_range+0x116/0x1a9 [2437548.269976] change_prot_numa+0x15/0x37 [2437548.274774] task_numa_work+0x1b8/0x302 [2437548.279512] task_work_run+0x62/0x95 [2437548.283882] exit_to_user_mode_loop+0x1a4/0x1a9 [2437548.289277] exit_to_user_mode_prepare+0xf4/0xfc [2437548.294751] ? sysvec_apic_timer_interrupt+0x34/0x81 [2437548.300677] irqentry_exit_to_user_mode+0x5/0x25 [2437548.306153] asm_sysvec_apic_timer_interrupt+0x16/0x1b
This is due to a race condition between change_prot_numa() and THP migration because the kernel doesn't check is_swap_pmd() and pmd_trans_huge() atomically:
change_prot_numa() THP migration ====================================================================== - change_pmd_range() -> is_swap_pmd() returns false, meaning it's not a PMD migration entry. - do_huge_pmd_numa_page() -> migrate_misplaced_page() sets migration entries for the THP. - change_pmd_range() -> pmd_none_or_clear_bad_unless_trans_huge() -> pmd_none() and pmd_trans_huge() returns false - pmd_none_or_clear_bad_unless_trans_huge() -> pmd_bad() returns true for the migration entry!
The upstream commit 670ddd8cdcbd ("mm/mprotect: delete pmd_none_or_clear_bad_unless_trans_huge()") closes this race condition by checking is_swap_pmd() and pmd_trans_huge() atomically.
# Backporting note
commit a79390f5d6a7 ("mm/mprotect: use long for page accountings and retval") is backported to return an error code (negative value) in change_pte_range().
Unlike the mainline, pte_offset_map_lock() does not check if the pmd entry is a migration entry or a hugepage; acquires PTL unconditionally instead of returning failure. Therefore, it is necessary to keep the !is_swap_pmd() && !pmd_trans_huge() && !pmd_devmap() checks in change_pmd_range() before acquiring the PTL.
After acquiring the lock, open-code the semantics of pte_offset_map_lock() in the mainline kernel; change_pte_range() fails if the pmd value has changed. This requires adding pmd_old parameter (pmd_t value that is read before calling the function) to change_pte_range().
Hugh Dickins (1): mm/mprotect: delete pmd_none_or_clear_bad_unless_trans_huge()
Peter Xu (1): mm/mprotect: use long for page accountings and retval
include/linux/hugetlb.h | 4 +- include/linux/mm.h | 2 +- mm/hugetlb.c | 4 +- mm/mempolicy.c | 2 +- mm/mprotect.c | 107 ++++++++++++++++++++++------------------ 5 files changed, 64 insertions(+), 55 deletions(-)