[Devel] Re: [RFC v14-rc3][PATCH 33/36] Support for share memory address spaces

[Oren Laadan]

On Thu, 9 Apr 2009, Serge E. Hallyn wrote:

> Quoting Oren Laadan (orenl at cs.columbia.edu):
> > 
> > 
> > Serge E. Hallyn wrote:
> > > Quoting Oren Laadan (orenl at cs.columbia.edu):
> > >> The task address space (task->mm) may be shared between processes if
> > >> CLONE_VM is used, and particularly among threads. Accordingly, treat
> > >> 'task->mm' as a shared object: during checkpoint check against the
> > >> objhash and only dump the contents if seen for the first time. During
> > >> restart, likewise, only restore if it's a new instance, otherwise use
> > >> the one already registered in the objhash.
> > >>
> > >> Signed-off-by: Oren Laadan <orenl at cs.columbia.edu>
> > > 
> > > Cool.
> > > 
> > > Acked-by: Serge Hallyn <serue at us.ibm.com>
> > > 
> > > Although:
> > > 
> > >> +	/* if the mm's objref is in the objhash, use that instance */
> > >> +	mm = cr_obj_get_by_ref(ctx, hh->objref, CR_OBJ_MM);
> > >> +	if (IS_ERR(mm)) {
> > >> +		ret = PTR_ERR(mm);
> > >> +		goto out;
> > >> +	}
> > >>
> > >> +	if (mm) {
> > >> +		if (mm != current->mm) {
> > > 
> > > In what twisted world could mm == current->mm at restart?
> > 
> > Tasks are re-created in user space, and so are threads. So threads will
> > already have their 'mm' set correctly.
> 
> Doesn't that assume that one task will complete sys_restart() before it
> does clone(CLONE_VM)?  Else sure, the threads will already share an mm,
> but it'll be the wrong one?  And I didn't think the sys_restart()
> synchronization supported that order.

During task creation, the algorithm implies that the thread group
leader is created first, and it in turn clones all the other threads
in the thread group. 

So now they all share the same MM, and no other task shares that mm.

One arbitrary thread is restarted first (depending on the checkpoint
order) - it will destory the VMAs in that MM and reconstruct new ones
within that MM.  When other threads get to cr_read_mm() they will
find the MM in the objhash and skip the reconstruction. Also, because
they already have the right MM, they will skip the re-attaching.

On the other hand, tasks that were cloned with VM_CLONE from any of
the threads in that thread group, will be created their own private
MM during restart, so in cr_read_mm() will need to really plug in
the MM found in the objhash. 

Oren.

> 
> (I realize I'm probably completely misunderstanding, and sounding like
> an idiot...)
> 
> And since OpenVZ has never re-sent their patch to do task creation in
> kernel-space on top of your set, I won't even debate about re-creation
> in user-space being certain  :)
> 
> -serge
> 
> 
_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers