Re: [PATCH] mm/khugepaged: Fix ->anon_vma race

["Kirill A. Shutemov" <kirill@shutemov.name>]

On Mon, Jan 16, 2023 at 02:07:41PM +0100, David Hildenbrand wrote:
> On 16.01.23 13:34, Kirill A. Shutemov wrote:
> > On Mon, Jan 16, 2023 at 01:06:59PM +0100, Jann Horn wrote:
> > > On Sun, Jan 15, 2023 at 8:07 PM Kirill A. Shutemov <kirill@shutemov.name> wrote:
> > > > On Fri, Jan 13, 2023 at 08:28:59PM +0100, Jann Horn wrote:
> > > > > No, that lockdep assert has to be there. Page table traversal is
> > > > > allowed under any one of the mmap lock, the anon_vma lock (if the VMA
> > > > > is associated with an anon_vma), and the mapping lock (if the VMA is
> > > > > associated with a mapping); and so to be able to remove page tables,
> > > > > we must hold all three of them.
> > > > 
> > > > Okay, that's fair. I agree with the patch now. Maybe adjust the commit
> > > > message a bit?
> > > 
> > > Just to make sure we're on the same page: Are you suggesting that I
> > > add this text?
> > > "Page table traversal is allowed under any one of the mmap lock, the
> > > anon_vma lock (if the VMA is associated with an anon_vma), and the
> > > mapping lock (if the VMA is associated with a mapping); and so to be
> > > able to remove page tables, we must hold all three of them."
> > > Or something else?
> > 
> > Looks good to me.
> > 
> > > > Anyway:
> > > > 
> > > > Acked-by: Kirill A. Shutemov <kirill.shutemov@intel.linux.com>
> > > 
> > > Thanks!
> > > 
> > > > BTW, I've noticied that you recently added tlb_remove_table_sync_one().
> > > > I'm not sure why it is needed. Why IPI in pmdp_collapse_flush() in not
> > > > good enough to serialize against GUP fast?
> > > 
> > > If that sent an IPI, it would be good enough; but
> > > pmdp_collapse_flush() is not guaranteed to send an IPI.
> > > It does a TLB flush, but on some architectures (including arm64 and
> > > also virtualized x86), a remote TLB flush can be done without an IPI.
> > > For example, arm64 has some fancy hardware support for remote TLB
> > > invalidation without IPIs ("broadcast TLB invalidation"), and
> > > virtualized x86 has (depending on the hypervisor) things like TLB
> > > shootdown hypercalls (under Hyper-V, see hyperv_flush_tlb_multi) or
> > > TLB shootdown signalling for preempted CPUs through shared memory
> > > (under KVM, see kvm_flush_tlb_multi).
> > 
> > I think such architectures must provide proper pmdp_collapse_flush()
> > with the required serialization. Power and S390 already do that.
> > 
> 
> The plan is to eventually move away from (ab)using IPI to synchronize with
> GUP-fast. Moving further into that direction a is wrong.
> 
> The flush was added as a quick fix for all architectures by Jann, until
> we can do better.
> 
> Even for ppc64, see:
> 
> commit bedf03416913d88c796288f9dca109a53608c745
> Author: Yang Shi <shy828301@gmail.com>
> Date:   Wed Sep 7 11:01:44 2022 -0700
> 
>     powerpc/64s/radix: don't need to broadcast IPI for radix pmd collapse flush
>     The IPI broadcast is used to serialize against fast-GUP, but fast-GUP will
>     move to use RCU instead of disabling local interrupts in fast-GUP.  Using
>     an IPI is the old-styled way of serializing against fast-GUP although it
>     still works as expected now.
>     And fast-GUP now fixed the potential race with THP collapse by checking
>     whether PMD is changed or not.  So IPI broadcast in radix pmd collapse
>     flush is not necessary anymore.  But it is still needed for hash TLB.

Okay. But I think tlb_remove_table_sync_one() belongs inside
pmdp_collapse_flush(). Collapsing pmd table into huge page without
serialization is a bug. They should not be separate.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov