{"id":"CVE-2026-97945","published":"2026-09-25T11:17:21.983","lastModified":"2026-09-25T11:17:21.983","description":"In the Linux kernel, the following vulnerability has been resolved:\n\nx86/mm: Fix user-space data loss with MADV_FREE and THP\n\nSome of users of Polars (a data analytics library) have lost production\ndata from this bug. They seem to have just the right combination of\nhuge pages, MADV_FREE and heavy reclaim pressure.\n\npmd_modify() masks the old value with (_HPAGE_CHG_MASK & ~_PAGE_DIRTY),\nsilently discarding the hardware dirty bit.  The subsequent\npmd_mksaveddirty() call is supposed to transfer _PAGE_DIRTY into\n_PAGE_SAVED_DIRTY when write-protecting, but the dirty bit was already\nstripped from the value, so there is nothing left to transfer.\n\nContrast with pte_modify(), which keeps _PAGE_DIRTY_BITS in its mask,\nand pud_modify(), which keeps _HPAGE_CHG_MASK untouched: pmd_modify()\nis the odd one out.  Any pmd_modify() on a writable, dirty PMD loses\nthe dirty state.\n\nOne visible consequence is data loss with MADV_FREE on PMD-mapped THP:\n\n    memset(buf, 0x5A, size);          // PMD-mapped THP, PMD dirty\n    madvise(buf, size, MADV_FREE);    // PMD cleaned but left writable,\n                                      // folio marked lazyfree\n    memset(buf, 0x5A, size);          // hardware sets _PAGE_DIRTY again\n    mprotect(buf, size, PROT_READ);   // pmd_modify() drops the dirty bit\n    mprotect(buf, size, PROT_READ|PROT_WRITE);\n    // ... memory pressure ...\n\nReclaim (e.g. under memcg pressure) then finds the lazyfree folio with\nno dirty bit set anywhere and frees it in\n__discard_anon_folio_pmd_locked(), even though the data was rewritten\nafter MADV_FREE; subsequent reads fault in fresh zero pages.  NUMA\nhinting alone can trigger the same loss, as do_huge_pmd_numa_page()\nrestores the PMD through pmd_modify() as well.\n\nPMD-mapped file THPs are affected too: mprotect()/NUMA hinting dropping\nthe dirty bit means rewritten data is never written back.\n\nFix it by keeping _PAGE_DIRTY in the preserved mask, exactly like\npte_modify() and pud_modify() do.  The existing\npmd_mksaveddirty()/pmd_clear_saveddirty() pair then performs the\nhardware-dirty <-> saved-dirty transition based on the write bit,\npreserving the shadow-stack encoding rules.","cvssScore":null,"cvssSeverity":null,"cvssVector":null,"cwes":[],"vendors":[],"products":[],"references":[{"url":"https://git.kernel.org/stable/c/55ddbc2ca6d5db64d44c3fb8f8705b4d39eb30b7","tags":[]},{"url":"https://git.kernel.org/stable/c/a5ad97cbcfcec31430268441c10eedde73200177","tags":[]},{"url":"https://git.kernel.org/stable/c/e4d569fe2f436a9b58c048fefabf56eab68c1b7e","tags":[]},{"url":"https://git.kernel.org/stable/c/f7491d7c81db0e7c304a7bd757a76d2fbeaff80e","tags":[]}],"exploitRefs":[],"hasPoc":false,"ai":null}