CVE-2026-72043

When hardware page table walker (PTW) is enabled on LoongArch, the CPU may set _PAGE_DIRTY directly in the page table entry during a write TLB miss, without going through the software TLB store handler. The software TLB store handler (tlbex.S:254) sets both _PAGE_DIRTY and_PAGE_MODIFIED together:

ori t0, t0, (_PAGE_VALID | _PAGE_DIRTY | _PAGE_MODIFIED)

Since hardware PTW only sets _PAGE_DIRTY, the software-only bit, i.e. _PAGE_MODIFIED is left unchanged. This creates a window where a PTE has _PAGE_DIRTY set (hardware knows the page is dirty) but _PAGE_MODIFIED clear (software is unaware).

When fork()/clone() triggers copy-on-write, __copy_present_ptes() calls pte_wrprotect(), which unconditionally clears both the _PAGE_WRITE and _PAGE_DIRTY bits:

pte_val(pte) &= ~(_PAGE_WRITE | _PAGE_DIRTY);

Since _PAGE_MODIFIED was never set, the dirtiness information is lost completely. Subsequently, when memory pressure triggers page reclaim, page_mkclean() / try_to_unmap() sees the page as clean (i.e. pte_dirty() returns false) and the page may be freed without writeback, causing data corruption.

Fix this by propagating the _PAGE_DIRTY bit to the _PAGE_MODIFIED bit in both pte_wrprotect() and pmd_wrprotect() before clearing writeable bits:

if (pte_val(pte) & _PAGE_DIRTY) pte_val(pte) |= _PAGE_MODIFIED;

The pmd_wrprotect() fix handles the CONFIG_TRANSPARENT_HUGEPAGE case, where pmd entries need the same treatment.

This ensures the software dirty tracking bit (checked by pte_dirty() and pmd_dirty(), which read both the _PAGE_DIRTY and _PAGE_MODIFIED bits) is preserved across fork COW write-protection.

The issue was found by the LTP madvise09 test case, which exercises page reclaim after "madvise(MADV_FREE), write and fork" operation sequence on private anonymous mappings.

No remediation recorded yet.