On Tue, 4 Nov 2014, Dave Hansen wrote:

On 11/03/2014 01:29 PM, Thomas Gleixner wrote:

quoted

On Mon, 3 Nov 2014, Dave Hansen wrote:

quoted

That's not really true. You can evaluate that information with
mmap_sem held for read as well. Nothing can change the mappings until
you drop it. So you could do:

   down_write(mm->bd_sem);
   down_read(mm->mmap_sem;
   evaluate_size_of_shm_to_unmap();
   clear_bounds_directory_entries();
   up_read(mm->mmap_sem);
   do_the_real_shm_unmap();
   up_write(mm->bd_sem);

That should still be covered by the above scheme.

Yep, that'll work.  It just means rewriting the shmdt()/mremap() code to
do a "dry run" of sorts.

Right. So either that or we hold bd_sem write locked accross all write
locked mmap_sem sections. Dunno, which solution is prettier :)

Do you have any concerns about adding another mutex to these paths?

You mean bd_sem? I don't think its an issue. You need to get mmap_sem
for write as well. So 

munmap() isn't as hot of a path as the allocation side, but it does
worry me a bit that we're going to perturb some workloads.  We might
need to find a way to optimize out the bd_sem activity on processes that
never used MPX.

I think using mm->bd_addr as a conditional for the bd_sem/mpx activity
is good enough. You just need to make sure that you store the result
of the starting conditional and use it for the closing one as well.

   mpx = mpx_pre_unmap(mm);
       {
	  if (!kernel_managing_bounds_tables(mm)
       	     return 0;
	  down_write(mm->bd_sem);
	  ...
	  return 1;
       }

   unmap();

   mxp_post_unmap(mm, mpx);
       {
          if (mpx) {
	     ....
	     up_write(mm->bd_sem);
       }

So this serializes nicely with the bd_sem protected write to
mm->bd_addr. There is a race there, but I don't think it matters. The
worst thing what can happen is a stale bound table.

Thanks,

	tglx