Re: [BUG?] X86 arch_tlbbatch_flush() seems to be lacking mm_tlb_flush_nested() integration

[Kirill A. Shutemov]

On Fri, Oct 14, 2022 at 08:19:42PM +0200, Jann Horn wrote:
> Hi!
> 
> I haven't actually managed to reproduce this behavior, so maybe I'm
> just misunderstanding how this works; but I think the
> arch_tlbbatch_flush() path for batched TLB flushing in vmscan ought to
> have some kind of integration with mm_tlb_flush_nested().
> 
> I think that currently, the following race could happen:
> 
> [initial situation: page P is mapped into a page table of task B, but
> the page is not referenced, the PTE's A/D bits are clear]
> A: vmscan begins
> A: vmscan looks at P and P's PTEs, and concludes that P is not currently in use
> B: reads from P through the PTE, setting the Accessed bit and creating
> a TLB entry
> A: vmscan enters try_to_unmap_one()
> A: try_to_unmap_one() calls should_defer_flush(), which returns true
> A: try_to_unmap_one() removes the PTE and queues a TLB flush
> (arch_tlbbatch_add_mm())
> A: try_to_unmap_one() returns, try_to_unmap() returns to shrink_folio_list()
> B: calls munmap() on the VMA that mapped P
> B: no PTEs are removed, so no TLB flush happens
> B: munmap() returns

I think here we will serialize against anon_vma/i_mmap lock in
__do_munmap() -> unmap_region() -> free_pgtables() that A also holds.

So I believe munmap() is safe, but MADV_DONTNEED (and its flavours) is not.

> [at this point, the TLB entry still exists]
> B: calls mmap(), which reuses the same area that was just unmapped
> B: tries to access the newly created VMA, but instead the access goes
> through the stale TLB entry
> A: shrink_folio_list() calls try_to_unmap_flush(), which removes the
> stale TLB entry
> 
> The effect would be that after process B removes a mapping with
> munmap() and creates a new mapping in its place, it would still see
> data from the old mapping when trying to access the new mapping.
> 
> Am I missing something that protects against this scenario?
> 
> munmap() uses the mmu_gather infrastructure, which tries to protect
> against this kind of correctness bug with multiple racing TLB
> invalidations in tlb_finish_mmu() by blowing away the whole TLB
> whenever one TLB invalidation ends while another is still in progress
> (tested with mm_tlb_flush_nested(tlb->mm)). But mmu_gather doesn't
> seem to be aware of TLB flushes that are batched up in the
> arch_tlbbatch_flush() infrastructure, so that doesn't help here.
> 
> I think it might be necessary to add a new global counter of pending
> arch_tlbbatch_flush() flushes, and query that in
> mm_tlb_flush_nested(), or something like that.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov