Date:	Tue, 7 Jul 2015 08:56:37 -0700
From:	Andy Lutomirski <luto@...capital.net>
To:	Andrew Vagin <avagin@...n.com>
Cc:	Andrey Vagin <avagin@...nvz.org>,
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
	Linux API <linux-api@...r.kernel.org>,
	Oleg Nesterov <oleg@...hat.com>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Cyrill Gorcunov <gorcunov@...nvz.org>,
	Pavel Emelyanov <xemul@...allels.com>,
	Roger Luethi <rl@...lgate.ch>, Arnd Bergmann <arnd@...db.de>,
	Arnaldo Carvalho de Melo <acme@...nel.org>,
	David Ahern <dsahern@...il.com>,
	Pavel Odintsov <pavel.odintsov@...il.com>
Subject: Re: [PATCH 0/24] kernel: add a netlink interface to get information
 about processes (v2)

On Tue, Jul 7, 2015 at 8:43 AM, Andrew Vagin <avagin@...n.com> wrote:
> On Mon, Jul 06, 2015 at 10:10:32AM -0700, Andy Lutomirski wrote:
>> On Mon, Jul 6, 2015 at 1:47 AM, Andrey Vagin <avagin@...nvz.org> wrote:
>> > Currently we use the proc file system, where all information are
>> > presented in text files, what is convenient for humans.  But if we need
>> > to get information about processes from code (e.g. in C), the procfs
>> > doesn't look so cool.
>> >
>> > From code we would prefer to get information in binary format and to be
>> > able to specify which information and for which tasks are required. Here
>> > is a new interface with all these features, which is called task_diag.
>> > In addition it's much faster than procfs.
>> >
>> > task_diag is based on netlink sockets and looks like socket-diag, which
>> > is used to get information about sockets.
>>
>> I think I like this in principle, but I have can see a few potential
>> problems with using netlink for this:
>>
>> 1. Netlink very naturally handles net namespaces, but it doesn't
>> naturally handle any other kind of namespace.  In fact, the taskstats
>> code that you're building on has highly broken user and pid namespace
>> support.  (Look for some obviously useless init_user_ns and
>> init_pid_ns references.  But that's only the obvious problem.  That
>> code calls current_user_ns() and task_active_pid_ns(current) from
>> .doit, which is, in turn, called from sys_write, and looking at
>> current's security state from sys_write is a big no-no.)
>>
>> You could partially fix it by looking at f_cred's namespaces, but that
>> would be a change of what it means to create a netlink socket, and I'm
>> not sure that's a good idea.
>
> If I don't miss something, all problems around pidns and userns are
> related with multicast functionality. task_diag is using
> request/response scheme and doesn't send multicast packets.

It has nothing to do with multicast.  task_diag needs to know what
pidns and userns to use for a request, but netlink isn't set up to
give you any reasonably way to do that.  A netlink socket is
fundamentally tied to a *net* ns (it's a socket, after all).  But you
can send it requests using write(2), and calling current_user_ns()
from write(2) is bad.  There's a long history of bugs and
vulnerabilities related to thinking that current_cred() and similar
are acceptable things to use in write(2) implementations.

>
>>
>> 2. These look like generally useful interfaces, which means that
>> people might want to use them in common non-system software, which
>> means that some of that software might get run inside of sandboxes
>> (Sandstorm, xdg-app, etc.)  Sandboxes like that might block netlink
>> outright, since it can't be usefully filtered by seccomp.  (This isn't
>> really the case now, since netlink route queries are too common, but
>> still.)
>>
>> 3. Netlink is a bit tedious to use from userspace.  Especially for
>> things like task_diag, which are really just queries that generate
>> single replies.
>
> I don't understand this point. Could you elaborate? I thought the
> netlink was designed for such purposes. (not only for them, but for them
> too)
>
> There are two features of netlink which are used.
>
> The netlink interface allows to split response into a few packets, if
> it's too big to be transferred for one iteration.
>

Netlink is fine for these use cases (if they were related to the
netns, not the pid ns or user ns), and it works.  It's still tedious
-- I bet that if you used a syscall, the user code would be
considerable shorter, though. :)

How would this be a problem if you used plain syscalls?  The user
would make a request, and the syscall would tell the user that their
result buffer was too small if it was, in fact, too small.

> And I want to mention "Memory mapped netlink I/O" functionality, which
> can be used to speed up task_diag.
>

IIRC memory-mapped netlink writes are terminally broken and therefore
neutered in current kernels (and hence no faster, and possibly slower,
than plain send(2)).  Memory-mapped reads are probably okay, but I
can't imagine that feature actually saving time in any real workload.
Almost all of the cpu time spent in task_diag will be in locking,
following pointers, formatting things, etc, and adding a memcpy will
almost certainly be lost in the noise.

--Andy
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists