Hi Laurent, Christophe, Michael, all,

On 11/3/20 5:11 PM, Laurent Dufour wrote:

Le 23/10/2020 à 14:28, Christophe Leroy a écrit :

[..]

quoted

quoted

quoted

That seems like it would work for CRIU and make sense in general?

Sorry for the late answer, yes this would make more sense.

Here is a patch doing that.

In your patch, the test seems overkill:

+    if ((start <= vdso_base && vdso_end <= end) ||  /* 1   */
+        (vdso_base <= start && start < vdso_end) || /* 3,4 */
+        (vdso_base < end && end <= vdso_end))       /* 2,3 */
+        mm->context.vdso_base = mm->context.vdso_end = 0;

What about

     if (start < vdso_end && vdso_start < end)
         mm->context.vdso_base = mm->context.vdso_end = 0;

This should cover all cases, or am I missing something ?


And do we really need to store vdso_end in the context ?
I think it should be possible to re-calculate it: the size of the VDSO
should be (&vdso32_end - &vdso32_start) + PAGE_SIZE for 32 bits VDSO,
and (&vdso64_end - &vdso64_start) + PAGE_SIZE for the 64 bits VDSO.

Thanks Christophe for the advise.

That is covering all the cases, and indeed is similar to the Michael's
proposal I missed last year.

I'll send a patch fixing this issue following your proposal.

It's probably not necessary anymore. I've sent patches [1], currently in
akpm, the last one forbids splitting of vm_special_mapping.
So, a user is able munmap() or mremap() vdso as a whole, but not partly.

[1]:
https://lore.kernel.org/linux-mm/20201013013416.390574-1-dima@arista.com/ (local)

Thanks,
          Dmitry