[PATCH v3 17/35] mm/mmap: write-lock VMA before shrinking or expanding it

[Liam R. Howlett]

* Suren Baghdasaryan <surenb at google.com> [230223 16:16]:
> On Thu, Feb 23, 2023 at 12:28 PM Liam R. Howlett
> <Liam.Howlett at oracle.com> wrote:
> >
> >
> > Wait, I figured a better place to do this.
> >
> > init_multi_vma_prep() should vma_start_write() on any VMA that is passed
> > in.. that we we catch any modifications here & in vma_merge(), which I
> > think is missed in this patch set?
> 
> Hmm. That looks like a good idea but in that case, why not do the
> locking inside vma_prepare() itself? From the description of that
> function it sounds like it was designed to acquire locks before VMA
> modifications, so would be the ideal location for doing that. WDYT?

That might be even better.  I think it will result in even less code.

There is also a vma_complete() which might work to call
vma_end_write_all() as well?

> The only concern is vma_adjust_trans_huge() being called before
> vma_prepare() but I *think* that's safe because
> vma_adjust_trans_huge() does its modifications after acquiring PTL
> lock, which page fault handlers also have to take. Does that sound
> right?

I am not sure.  We are certainly safe the way it is, and the PTL has to
be safe for concurrent faults.. but this could alter the walk to a page
table while that walk is occurring and I don't think that happens today.

It might be best to leave the locking order the way you have it, unless
someone can tell us it's safe?

We could pass through the three extra variables that are needed to move
the vma_adjust_trans_huge() call within that function as well?  This
would have the added benefit of having all locking grouped in the one
location, but the argument list would be getting long, however we could
use the struct.

remove & remove2 should be be detached in vma_prepare() or
vma_complete() as well?

> 
> >
> >
> > * Liam R. Howlett <Liam.Howlett at Oracle.com> [230223 15:20]:
> > > Reviewed-by: Liam R. Howlett <Liam.Howlett at oracle.com>
> > >
> > > * Suren Baghdasaryan <surenb at google.com> [230216 00:18]:
> > > > vma_expand and vma_shrink change VMA boundaries. Expansion might also
> > > > result in freeing of an adjacent VMA. Write-lock affected VMAs to prevent
> > > > concurrent page faults.
> > > >
> > > > Signed-off-by: Suren Baghdasaryan <surenb at google.com>
> > > > ---
> > > >  mm/mmap.c | 5 +++++
> > > >  1 file changed, 5 insertions(+)
> > > >
> > > > diff --git a/mm/mmap.c b/mm/mmap.c
> > > > index ec2f8d0af280..f079e5bbcd57 100644
> > > > --- a/mm/mmap.c
> > > > +++ b/mm/mmap.c
> > > > @@ -674,6 +674,9 @@ int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma,
> > > >             ret = dup_anon_vma(vma, next);
> > > >             if (ret)
> > > >                     return ret;
> > > > +
> > > > +           /* Lock the VMA  before removing it */
> > > > +           vma_start_write(next);
> > > >     }
> > > >
> > > >     init_multi_vma_prep(&vp, vma, NULL, remove_next ? next : NULL, NULL);
> > > > @@ -686,6 +689,7 @@ int vma_expand(struct vma_iterator *vmi, struct vm_area_struct *vma,
> > > >     if (vma_iter_prealloc(vmi))
> > > >             goto nomem;
> > > >
> > > > +   vma_start_write(vma);
> > > >     vma_adjust_trans_huge(vma, start, end, 0);
> > > >     /* VMA iterator points to previous, so set to start if necessary */
> > > >     if (vma_iter_addr(vmi) != start)
> > > > @@ -725,6 +729,7 @@ int vma_shrink(struct vma_iterator *vmi, struct vm_area_struct *vma,
> > > >     if (vma_iter_prealloc(vmi))
> > > >             return -ENOMEM;
> > > >
> > > > +   vma_start_write(vma);
> > > >     init_vma_prep(&vp, vma);
> > > >     vma_adjust_trans_huge(vma, start, end, 0);
> > > >     vma_prepare(&vp);
> > > > --
> > > > 2.39.1
> > > >
> >
> > --
> > To unsubscribe from this group and stop receiving emails from it, send an email to kernel-team+unsubscribe at android.com.
> >