On Wed, 22 Jan 2025, Roman Gushchin wrote:
> Commit b67fbebd4cf9 ("mmu_gather: Force tlb-flush VM_PFNMAP vmas")
> added a forced tlbflush to tlb_vma_end(), which is required to avoid a
> race between munmap() and unmap_mapping_range(). However it added some
> overhead to other paths where tlb_vma_end() is used, but vmas are not
> removed, e.g. madvise(MADV_DONTNEED).
>
> Fix this by moving the tlb flush out of tlb_end_vma() into
> free_pgtables(), somewhat similar to the stable version of the
> original commit: e.g. stable commit 895428ee124a ("mm: Force TLB flush
> for PFNMAP mappings before unlink_file_vma()").
>
> Note, that if tlb->fullmm is set, no flush is required, as the whole
> mm is about to be destroyed.
>
> v2:
> - moved vma_pfn flag handling into tlb.h (by Peter Z.)
> - added comments (by Peter Z.)
> - fixed the vma_pfn flag setting (by Hugh D.)
>
> Suggested-by: Jann Horn <jannh@xxxxxxxxxx>
> Signed-off-by: Roman Gushchin <roman.gushchin@xxxxxxxxx>
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Cc: Will Deacon <will@xxxxxxxxxx>
> Cc: "Aneesh Kumar K.V" <aneesh.kumar@xxxxxxxxxx>
> Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> Cc: Nick Piggin <npiggin@xxxxxxxxx>
> Cc: Hugh Dickins <hughd@xxxxxxxxxx>
> Cc: linux-arch@xxxxxxxxxxxxxxx
> Cc: linux-mm@xxxxxxxxx
> ---
> include/asm-generic/tlb.h | 41 ++++++++++++++++++++++++++-------------
> mm/memory.c | 7 +++++++
> 2 files changed, 35 insertions(+), 13 deletions(-)
>
> diff --git a/include/asm-generic/tlb.h b/include/asm-generic/tlb.h
> index 709830274b75..fbe31f49a5af 100644
> --- a/include/asm-generic/tlb.h
> +++ b/include/asm-generic/tlb.h
> @@ -449,7 +449,14 @@ tlb_update_vma_flags(struct mmu_gather *tlb, struct vm_area_struct *vma)
> */
> tlb->vma_huge = is_vm_hugetlb_page(vma);
> tlb->vma_exec = !!(vma->vm_flags & VM_EXEC);
> - tlb->vma_pfn = !!(vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP));
> +
> + /*
> + * vma_pfn is checked and cleared by tlb_flush_mmu_pfnmap()
> + * for a set of vma's, so it should be set if at least one vma
> + * has VM_PFNMAP or VM_MIXEDMAP flags set.
> + */
> + if (vma->vm_flags & (VM_PFNMAP|VM_MIXEDMAP))
> + tlb->vma_pfn = 1;
Okay, but struct mmu_gather is usually on the caller's stack,
containing junk initially, and there's nothing in this patch
yet to initialize tlb->vma_pfn to 0.
__tlb_reset_range() needs to do that, doesn't it? With some adjustment
to its comment about not resetting mmu_gather::vma_* fields. Or would
it be better to get around that by renaming vma_pfn to, er, something
else - I'd have to understand the essence of Jann's race better to
come up with the right name - untracked_mappings? would that be right?
I still haven't grasped the essence of that race.
(Panic attack: where is, for example, tlb->need_flush_all initialized
to 0? Ah, over in mm/mmu_gather.c, __tlb_gather_mmu(). Phew.)
And if __tlb_reset_range() resets tlb->vma_pfn to 0, then that has the
side-effect that any TLB flush cancels the vma_pfn state: which is a
desirable side-effect, isn't it? avoiding the possibility of doing an
unnecessary extra TLB flush in free_pgtables(), as I criticized before.
Hugh
> }
>
> static inline void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
> @@ -466,6 +473,22 @@ static inline void tlb_flush_mmu_tlbonly(struct mmu_gather *tlb)
> __tlb_reset_range(tlb);
> }
>
> +static inline void tlb_flush_mmu_pfnmap(struct mmu_gather *tlb)
> +{
> + /*
> + * VM_PFNMAP and VM_MIXEDMAP maps are fragile because the core mm
> + * doesn't track the page mapcount -- there might not be page-frames
> + * for these PFNs after all. Force flush TLBs for such ranges to avoid
> + * munmap() vs unmap_mapping_range() races.
> + * Ensure we have no stale TLB entries by the time this mapping is
> + * removed from the rmap.
> + */
> + if (unlikely(!tlb->fullmm && tlb->vma_pfn)) {
> + tlb_flush_mmu_tlbonly(tlb);
> + tlb->vma_pfn = 0;
> + }
> +}
> +
> static inline void tlb_remove_page_size(struct mmu_gather *tlb,
> struct page *page, int page_size)
> {
> @@ -549,22 +572,14 @@ static inline void tlb_start_vma(struct mmu_gather *tlb, struct vm_area_struct *
>
> static inline void tlb_end_vma(struct mmu_gather *tlb, struct vm_area_struct *vma)
> {
> - if (tlb->fullmm)
> + if (tlb->fullmm || IS_ENABLED(CONFIG_MMU_GATHER_MERGE_VMAS))
> return;
>
> /*
> - * VM_PFNMAP is more fragile because the core mm will not track the
> - * page mapcount -- there might not be page-frames for these PFNs after
> - * all. Force flush TLBs for such ranges to avoid munmap() vs
> - * unmap_mapping_range() races.
> + * Do a TLB flush and reset the range at VMA boundaries; this avoids
> + * the ranges growing with the unused space between consecutive VMAs.
> */
> - if (tlb->vma_pfn || !IS_ENABLED(CONFIG_MMU_GATHER_MERGE_VMAS)) {
> - /*
> - * Do a TLB flush and reset the range at VMA boundaries; this avoids
> - * the ranges growing with the unused space between consecutive VMAs.
> - */
> - tlb_flush_mmu_tlbonly(tlb);
> - }
> + tlb_flush_mmu_tlbonly(tlb);
> }
>
> /*
> diff --git a/mm/memory.c b/mm/memory.c
> index 398c031be9ba..c2a9effb2e32 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -365,6 +365,13 @@ void free_pgtables(struct mmu_gather *tlb, struct ma_state *mas,
> {
> struct unlink_vma_file_batch vb;
>
> + /*
> + * VM_PFNMAP and VM_MIXEDMAP maps require a special handling here:
> + * force flush TLBs for such ranges to avoid munmap() vs
> + * unmap_mapping_range() races.
> + */
> + tlb_flush_mmu_pfnmap(tlb);
> +
> do {
> unsigned long addr = vma->vm_start;
> struct vm_area_struct *next;
> --
> 2.48.1.262.g85cc9f2d1e-goog
