[PATCH v3 1/7] kernel/fork: convert vma assignment to a memcpy

[Marco Elver]

On Thu, Jan 26, 2023 at 09:27AM -0800, Paul E. McKenney wrote:
> On Wed, Jan 25, 2023 at 05:34:49PM -0800, Andrew Morton wrote:
> > On Wed, 25 Jan 2023 16:50:01 -0800 Suren Baghdasaryan <surenb at google.com> wrote:
> > 
> > > On Wed, Jan 25, 2023 at 4:22 PM Andrew Morton <akpm at linux-foundation.org> wrote:
> > > >
> > > > On Wed, 25 Jan 2023 15:35:48 -0800 Suren Baghdasaryan <surenb at google.com> wrote:
> > > >
> > > > > Convert vma assignment in vm_area_dup() to a memcpy() to prevent compiler
> > > > > errors when we add a const modifier to vma->vm_flags.
> > > > >
> > > > > ...
> > > > >
> > > > > --- a/kernel/fork.c
> > > > > +++ b/kernel/fork.c
> > > > > @@ -482,7 +482,7 @@ struct vm_area_struct *vm_area_dup(struct vm_area_struct *orig)
> > > > >                * orig->shared.rb may be modified concurrently, but the clone
> > > > >                * will be reinitialized.
> > > > >                */
> > > > > -             *new = data_race(*orig);
> > > > > +             memcpy(new, orig, sizeof(*new));
> > > >
> > > > The data_race() removal is unchangelogged?
> > > 
> > > True. I'll add a note in the changelog about that. Ideally I would
> > > like to preserve it but I could not find a way to do that.
> > 
> > Perhaps Paul can comment?
> > 
> > I wonder if KCSAN knows how to detect this race, given that it's now in
> > a memcpy.  I assume so.
> 
> I ran an experiment memcpy()ing between a static array and an onstack
> array, and KCSAN did not complain.  But maybe I was setting it up wrong.
> 
> This is what I did:
> 
> 	long myid = (long)arg; /* different value for each task */
> 	static unsigned long z1[10] = { 0 };
> 	unsigned long z2[10];
> 
> 	...
> 
> 	memcpy(z1, z2, ARRAY_SIZE(z1) * sizeof(z1[0]));
> 	for (zi = 0; zi < ARRAY_SIZE(z1); zi++)
> 		z2[zi] += myid;
> 	memcpy(z2, z1, ARRAY_SIZE(z1) * sizeof(z1[0]));
> 
> Adding Marco on CC for his thoughts.

( Sorry for not seeing it earlier - just saw this by chance. )

memcpy() data races will be detected as of (given a relatively recent
Clang compiler):

  https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=7c201739beef

Also beware that the compiler is free to "optimize" things by either
inlining memcpy() (turning an explicit memcpy() into just a bunch of
loads/stores), or outline plain assignments into memcpy() calls. So the
only way to be sure what ends up there is to look at the disassembled
code.

The data_race() was introduced by:

  https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=cda099b37d716

It says:
 "vm_area_dup() blindly copies all fields of original VMA to the new one.
  This includes coping vm_area_struct::shared.rb which is normally
  protected by i_mmap_lock. But this is fine because the read value will
  be overwritten on the following __vma_link_file() under proper
  protection. Thus, mark it as an intentional data race and insert a few
  assertions for the fields that should not be modified concurrently."

And as far as I can tell this hasn't changed.

Thanks,
-- Marco