cryo and mm->arg_start
matthltc at us.ibm.com
Fri Jul 11 15:01:13 PDT 2008
On Fri, 2008-07-11 at 09:38 -0700, Dave Hansen wrote:
> On Fri, 2008-07-11 at 08:13 -0500, Serge E. Hallyn wrote:
> > One thing we could do here is to start extending the cryo approach
> > with Eric's checkpoint-as-a-coredump (caac?). We generate the
> > tiniest of coredumps which, at first, contains nothing but
> > mm->arg_start and maybe a process id. It would be simplest if
> > it also contained a filename for the real executable,
> The exec model sounds reasonable to me.
> But, I think the filename of the exe is going to have to be in the
> checkpoint *already*. It is mapped by at least one of the VMAs, and
> will probably be dumped as a normal file-backed area.
Yes, the file that backed the exec will be there. Note that thanks to
"stacking" filesystems the path to the file backing the exe is not
_always_ going to be the same as the path to the file which userspace
exec'd in the first place. You can see this by comparing
the /proc/<pid>/exe symlink with the file backing the VMA.
This is important to any program which checks the /proc/self/exe
symlink to find out where it's installed (Java does this, for example).
I think it's possible to do this with a binfmt -- it's just one more
detail to remember.
More information about the Containers