On Mon, Dec 17, 2012 at 05:04:08PM +0100, Oleg Nesterov wrote:
> On 12/17, Neil Horman wrote:
> >
> > On Mon, Dec 17, 2012 at 01:34:28PM +0100, Oleg Nesterov wrote:
> > > @@ -455,6 +468,14 @@ static int umh_pipe_setup(struct subproc
> > > /* and disallow core files too */
> > > current->signal->rlim[RLIMIT_CORE] = (struct rlimit){1, 1};
> > >
> > > +
> > > + if (cp->switch_ns) {
> > > + get_fs_root(cp->cprocess->fs, &root);
> > > + set_fs_root(current->fs, &root);
> > > + switch_task_namespaces(current, cp->cprocess->nsproxy);
> > >
> > > How? You can't simply change ->nsproxy this way.
> > >
> > Why not? This is exactly how fork, exit, and setns use this call.
>
> No. exit() does switch_task_namespaces(NULL), this is different.
> fork() doesn't do this, and unshare/setns carefully creates the new ns.
>
> > > If nothing else this breaks sys_getpid(), no?
> > >
> > hmm, I think you're inferring here that there is a chance that a pid
> > allocated
> > in the init namespace might conflict with another process who holds the
> > same pid
> > in another namespace?
>
> No, I meant that sys_getpid() should always return 0 after this
> switch_task_namespaces() if the coredumping task is not from the root
> namespace.
>
> > Is there a way to switch all namespaces, except for the pid
> > namespace?
>
> Which exactly namespaces you want to change?
>
Ideally, I want the pipe reader process to execute in the same namespaces that
the crashing process executed in (i.e. the pipe reader should execute as though
the crashing process forked it).
> To be honest, I do not understand this patch at all. It seems that
> you need to do something like sys_setns(). But if we do this, then
> why we can't make core_pattern per-namespace?
>
That actually would make sense, although we can't really use setns directly, as
I don't think we want to open file descriptors to do this manipulation in the
kernel.
The origional goal of this patch was to allow for the core_pattern pipe reader
to execute using the same root fs point that the crashing process did (the idea
being that, if a process crashed, the pipe reader should execute in the same
environment. E.g. if a container was running a process that crashed, and the
pipe reader was going to save the core to /tmp, it should be in the chrooted
/tmp, not the global /tmp. But it occured to me that switching all the
namespaces is really whats needed here. Although after this conversation, that
seems far more difficult than I origionally thought.
Perhaps its best just to restrict this patch to adjusting the root fs location
for the chroot case. Or would it be better to iterate over the setns-able
namespaces and migrate the pipe helper to each of them from umh_pipe_setup.
Thaoughts?
Neil
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
