Messages in this thread |  | | Subject | Re: [RFC PATCH 3/8] kmod - teach call_usermodehelper() to use a namespace | From | Ian Kent <> | Date | Wed, 18 Feb 2015 09:42:25 +0800 |
| |
On Mon, 2015-02-16 at 18:13 +0100, Oleg Nesterov wrote: > On 02/16, Ian Kent wrote: > > > > On Tue, 2015-02-10 at 17:55 +0100, Oleg Nesterov wrote: > > > On 02/10, Ian Kent wrote: > > > > > > > > On Mon, 2015-02-09 at 17:03 +0100, Oleg Nesterov wrote: > > > > > > > > > > I understand. but I still can't understand why we can't implement something > > > > > like > > > > > enter_ns(struct nsproxy *p) > > > > > { > > > > > new_nsproxy = create_new_namespaces(...); > > > > > > > > > > p->mnt_ns->ns->ops->install(new_nsproxy, ...); > > > > > p->pid_ns_for_children->ns->ops->install(new_nsproxy, ...); > > > > > ... > > > > > > > > > > switch_task_namespaces(new_nsproxy); > > > > > } > > > > > > > > > > Why we should abuse fs/proc ? > > > > > > > > That sounds like a much better approach. > > > > Your saying just take a reference to the nsproxy from the located > > > > process and use it instead, right? > > > > > > Yes, > > > > I'm still not sure if this can be done (at least without surgery to the > > namespace implementation) and I think I've been here before which is > > what lead to the file_open_root() approach. > > > > The difficulty is the second parameter to the install() call above, the > > struct ns_common. In setns() it's obtained from the procfs file inode > > and the file open is what's used to obtain each namespace type (in the > > form of a struct ns_common) from a process context different from > > current, current being the thread runner process. > > > > I might still be able to work out a (viable) way to obtain the > > appropriate ns_common struct in each case without a file open but it's > > hard to see how. > > Not sure I understand... Every "namespace" pointer "struct nsproxy" has > the "struct ns_common ns" you need? So you can do (for example) > > p->mnt_ns->ns->ops->install(new_nsproxy, &p->mnt_ns->ns); > > or I missed something? (userns differs, you need cred->user_ns, of course).
I didn't see that when I looked so I missed it, thanks for pointing it out.
> > > Perhaps I missed something, but this doesn't look like a problem... > > The real problem is that , let me repeat, is that pidns_install() does not > and can't change active_pid_ns. So I think that kernel_thread_in_ns() probably > make more sense.
Right, I didn't miss that when you mentioned it.
Changing the execution order using your kernel_thread_in_ns() suggestion is clearly what needs to be done.
Ian
|  |