On 23/01/2025 14:38, Ryan Roberts wrote: > I think there might be a bug in this after all... > > > On 07/01/2025 14:47, Ryan Roberts wrote: >> When mremap()ing a memory region previously registered with userfaultfd >> as write-protected but without UFFD_FEATURE_EVENT_REMAP, an >> inconsistency in flag clearing leads to a mismatch between the vma flags >> (which have uffd-wp cleared) and the pte/pmd flags (which do not have >> uffd-wp cleared). This mismatch causes a subsequent mprotect(PROT_WRITE) >> to trigger a warning in page_table_check_pte_flags() due to setting the >> pte to writable while uffd-wp is still set. >> >> Fix this by always explicitly clearing the uffd-wp pte/pmd flags on any >> such mremap() so that the values are consistent with the existing >> clearing of VM_UFFD_WP. Be careful to clear the logical flag regardless >> of its physical form; a PTE bit, a swap PTE bit, or a PTE marker. Cover >> PTE, huge PMD and hugetlb paths. >> >> Co-developed-by: Mikołaj Lenczewski <miko.lenczewski@xxxxxxx> >> Signed-off-by: Mikołaj Lenczewski <miko.lenczewski@xxxxxxx> >> Signed-off-by: Ryan Roberts <ryan.roberts@xxxxxxx> >> Closes: https://lore.kernel.org/linux-mm/810b44a8-d2ae-4107-b665-5a42eae2d948@xxxxxxx/ >> Fixes: 63b2d4174c4a ("userfaultfd: wp: add the writeprotect API to userfaultfd ioctl") >> Cc: stable@xxxxxxxxxxxxxxx >> --- >> include/linux/userfaultfd_k.h | 12 ++++++++++++ >> mm/huge_memory.c | 12 ++++++++++++ >> mm/hugetlb.c | 14 +++++++++++++- >> mm/mremap.c | 32 +++++++++++++++++++++++++++++++- >> 4 files changed, 68 insertions(+), 2 deletions(-) >> >> diff --git a/include/linux/userfaultfd_k.h b/include/linux/userfaultfd_k.h >> index cb40f1a1d081..75342022d144 100644 >> --- a/include/linux/userfaultfd_k.h >> +++ b/include/linux/userfaultfd_k.h >> @@ -247,6 +247,13 @@ static inline bool vma_can_userfault(struct vm_area_struct *vma, >> vma_is_shmem(vma); >> } >> >> +static inline bool vma_has_uffd_without_event_remap(struct vm_area_struct *vma) >> +{ >> + struct userfaultfd_ctx *uffd_ctx = vma->vm_userfaultfd_ctx.ctx; >> + >> + return uffd_ctx && (uffd_ctx->features & UFFD_FEATURE_EVENT_REMAP) == 0; >> +} >> + >> extern int dup_userfaultfd(struct vm_area_struct *, struct list_head *); >> extern void dup_userfaultfd_complete(struct list_head *); >> void dup_userfaultfd_fail(struct list_head *); >> @@ -402,6 +409,11 @@ static inline bool userfaultfd_wp_async(struct vm_area_struct *vma) >> return false; >> } >> >> +static inline bool vma_has_uffd_without_event_remap(struct vm_area_struct *vma) >> +{ >> + return false; >> +} >> + >> #endif /* CONFIG_USERFAULTFD */ >> >> static inline bool userfaultfd_wp_use_markers(struct vm_area_struct *vma) >> diff --git a/mm/huge_memory.c b/mm/huge_memory.c >> index c89aed1510f1..2654a9548749 100644 >> --- a/mm/huge_memory.c >> +++ b/mm/huge_memory.c >> @@ -2212,6 +2212,16 @@ static pmd_t move_soft_dirty_pmd(pmd_t pmd) >> return pmd; >> } >> >> +static pmd_t clear_uffd_wp_pmd(pmd_t pmd) >> +{ >> + if (pmd_present(pmd)) >> + pmd = pmd_clear_uffd_wp(pmd); >> + else if (is_swap_pmd(pmd)) >> + pmd = pmd_swp_clear_uffd_wp(pmd); >> + >> + return pmd; >> +} >> + >> bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, >> unsigned long new_addr, pmd_t *old_pmd, pmd_t *new_pmd) >> { >> @@ -2250,6 +2260,8 @@ bool move_huge_pmd(struct vm_area_struct *vma, unsigned long old_addr, >> pgtable_trans_huge_deposit(mm, new_pmd, pgtable); >> } >> pmd = move_soft_dirty_pmd(pmd); >> + if (vma_has_uffd_without_event_remap(vma)) >> + pmd = clear_uffd_wp_pmd(pmd); >> set_pmd_at(mm, new_addr, new_pmd, pmd); >> if (force_flush) >> flush_pmd_tlb_range(vma, old_addr, old_addr + PMD_SIZE); >> diff --git a/mm/hugetlb.c b/mm/hugetlb.c >> index 354eec6f7e84..cdbc55d5384f 100644 >> --- a/mm/hugetlb.c >> +++ b/mm/hugetlb.c >> @@ -5454,6 +5454,7 @@ static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, >> unsigned long new_addr, pte_t *src_pte, pte_t *dst_pte, >> unsigned long sz) >> { >> + bool need_clear_uffd_wp = vma_has_uffd_without_event_remap(vma); >> struct hstate *h = hstate_vma(vma); >> struct mm_struct *mm = vma->vm_mm; >> spinlock_t *src_ptl, *dst_ptl; >> @@ -5470,7 +5471,18 @@ static void move_huge_pte(struct vm_area_struct *vma, unsigned long old_addr, >> spin_lock_nested(src_ptl, SINGLE_DEPTH_NESTING); >> >> pte = huge_ptep_get_and_clear(mm, old_addr, src_pte); >> - set_huge_pte_at(mm, new_addr, dst_pte, pte, sz); >> + >> + if (need_clear_uffd_wp && pte_marker_uffd_wp(pte)) >> + huge_pte_clear(mm, new_addr, dst_pte, sz); > > This is checking if the source huge_pte is a uffd-wp marker and clearing the > destination if so. The destination could have previously held arbitrary valid > mappings, I guess? > > But huge_pte_clear() does not call page_table_check_pte_clear(). So any previous > valid mapping will not have it's page_table_check ref count decremented? > > I think it should be replaced with: > huge_ptep_get_and_clear(mm, new_addr, dst_pte); > > Since there is no huge_ptep_clear(). > > The tests I wrote are always mremapping into PROT_NONE space so they don't hit > this condition. If people agree this is a bug, I'll send out a fix. OK I'm deep in the rabbit hole now; Could anyone clarify the specs for huge_pte_clear() and huge_ptep_get_and_clear()? Specifically, I believe: - huge_pte_clear() can only be called for not-present huge_ptes, because the only way to set a huge_pte is via set_huge_pte_at() and that will always execute the page_table_check_*_set() functions for present huge_ptes. So clearing a present huge_pte using huge_pte_clear(), which does not call page_table_check_*_clear(), would cause counter imbalance. It looks like existing generic-code callers of huge_pte_clear() only do it for non-present huge_ptes. - huge_ptep_get_and_clear() can be used to get-and-clear both present and non-present huge_ptes? There is a single call site in generic-code where this could be called for a non-present huge_pte: move_huge_pte() (and it was there prior to my change). BUT, it looks like the arm64 implementation of huge_ptep_get_and_clear() is not safe to call if the pte being cleared is non-present: pte_t huge_ptep_get_and_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep) { int ncontig; size_t pgsize; pte_t orig_pte = __ptep_get(ptep); if (!pte_cont(orig_pte)) return __ptep_get_and_clear(mm, addr, ptep); ncontig = find_num_contig(mm, addr, ptep, &pgsize); return get_clear_contig(mm, addr, ptep, pgsize, ncontig); } pte_cont() assumes the pte is present, otherwise it's sampling a random bit that doesn't have the meaning it thinks it has. It is currently relying on that to determine the size of the huge_pte. So either arm64 has a bug or move_huge_pte() has a bug. If huge_ptep_get_and_clear() needs to work for non-present huge_ptes, we will need to pass the huge_pte size into this function. I don't think there is another way to resolve this. Thanks, Ryan > > Thanks, > Ryan >
