RE: [PATCH v2 3/3] Make core_pattern support namespace

[Zhao Lei]

Hi, Eric

> -----Original Message-----
> From: Eric W. Biederman [mailto:ebiederm@xxxxxxxxxxxx]
> Sent: Monday, March 21, 2016 4:15 PM
> To: Zhao Lei <zhaolei@xxxxxxxxxxxxxx>
> Cc: linux-kernel@xxxxxxxxxxxxxxx; containers@xxxxxxxxxxxxxxxxxxxxxxxxxx;
> 'Mateusz Guzik' <mguzik@xxxxxxxxxx>; 'Kamezawa Hiroyuki'
> <kamezawa.hiroyu@xxxxxxxxxxxxxx>
> Subject: Re: [PATCH v2 3/3] Make core_pattern support namespace
> 
> Zhao Lei <zhaolei@xxxxxxxxxxxxxx> writes:
> 
> > Hi, Eric W. Biederman
> >
> >> -----Original Message-----
> >> From: Eric W. Biederman [mailto:ebiederm@xxxxxxxxxxxx]
> >> Sent: Monday, March 21, 2016 2:00 PM
> >> To: Zhao Lei <zhaolei@xxxxxxxxxxxxxx>
> >> Cc: linux-kernel@xxxxxxxxxxxxxxx; containers@xxxxxxxxxxxxxxxxxxxxxxxxxx;
> >> Mateusz Guzik <mguzik@xxxxxxxxxx>
> >> Subject: Re: [PATCH v2 3/3] Make core_pattern support namespace
> >>
> >> Zhao Lei <zhaolei@xxxxxxxxxxxxxx> writes:
> >>
> >> > Currently, each container shared one copy of coredump setting
> >> > with the host system, if host system changed the setting, each
> >> > running containers will be affected.
> >> >
> >> > Moreover, it is not easy to let each container keeping their own
> >> > coredump setting.
> >> >
> >> > We can use some workaround as pipe program to make the second
> >> > requirement possible, but it is not simple, and both host and
> >> > container are limited to set to fixed pipe program.
> >> > In one word, for host running contailer, we can't change core_pattern
> >> > anymore.
> >> > To make the problem more hard, if a host running more than one
> >> > container product, each product will try to snatch the global
> >> > coredump setting to fit their own requirement.
> >> >
> >> > For container based on namespace design, it is good to allow
> >> > each container keeping their own coredump setting.
> >> >
> >> > It will bring us following benefit:
> >> > 1: Each container can change their own coredump setting
> >> >    based on operation on /proc/sys/kernel/core_pattern
> >> > 2: Coredump setting changed in host will not affect
> >> >    running containers.
> >> > 3: Support both case of "putting coredump in guest" and
> >> >    "putting curedump in host".
> >> >
> >> > Each namespace-based software(lxc, docker, ..) can use this function
> >> > to custom their dump setting.
> >> >
> >> > And this function makes each continer working as separate system,
> >> > it fit for design goal of namespace
> >>
> >> There are a lot of questionable things with this patchset.
> >>
> >> > @@ -183,7 +182,7 @@ put_exe_file:
> >> >  static int format_corename(struct core_name *cn, struct
> >> coredump_params *cprm)
> >> >  {
> >> >  	const struct cred *cred = current_cred();
> >> > -	const char *pat_ptr = core_pattern;
> >> > +	const char *pat_ptr =
> >> current->nsproxy->pid_ns_for_children->core_pattern;
> >>
> >> current->nsproxy->pid_ns_for_children as the name implies is completely
> >> inappropriate for getting the pid namespace of the current task.
> >>
> >> This should use task_active_pid_namespace.
> >>
> > In 5 members in nsproxy struct, pid_ns_for_children seems the best place
> > for this variable.
> 
> nsproxy is not a magic list of namespaces, nsproxy is to keep
> task_struct from expanding and as a trick to keep reference count
> increments for namespaces cheap.
> 
> > And no variable named task_active_pid_namespace in source,
> > could you explain it deeply?
> 
> Apologies I mispelled it.  Look in pid_namespace.h at
> task_active_pid_ns.  If you want a tasks pid namespace that is the
> function to use.
> 
> pid_ns_for_children only describes newly forked children.  Which leads
> to another problem of your patchset.  I can force your coredump helper
> into a pid namespace that the program that dumps core will create it's
> children in.
> 
> >> >  	int ispipe = (*pat_ptr == '|');
> >> >  	int pid_in_pattern = 0;
> >> >  	int err = 0;
> >> > diff --git a/include/linux/pid_namespace.h
> b/include/linux/pid_namespace.h
> >> > index 918b117..a5af1e9 100644
> >> > --- a/include/linux/pid_namespace.h
> >> > +++ b/include/linux/pid_namespace.h
> >> > @@ -9,6 +9,7 @@
> >> >  #include <linux/nsproxy.h>
> >> >  #include <linux/kref.h>
> >> >  #include <linux/ns_common.h>
> >> > +#include <linux/binfmts.h>
> >> >
> >> >  struct pidmap {
> >> >         atomic_t nr_free;
> >> > @@ -45,6 +46,7 @@ struct pid_namespace {
> >> >  	int hide_pid;
> >> >  	int reboot;	/* group exit code if this pidns was rebooted */
> >> >  	struct ns_common ns;
> >> > +	char core_pattern[CORENAME_MAX_SIZE];
> >> >  };
> >> >
> >> >  extern struct pid_namespace init_pid_ns;
> >> > diff --git a/kernel/pid.c b/kernel/pid.c
> >> > index 4d73a83..c79c1d5 100644
> >> > --- a/kernel/pid.c
> >> > +++ b/kernel/pid.c
> >> > @@ -83,6 +83,7 @@ struct pid_namespace init_pid_ns = {
> >> >  #ifdef CONFIG_PID_NS
> >> >  	.ns.ops = &pidns_operations,
> >> >  #endif
> >> > +	.core_pattern = "core",
> >> >  };
> >> >  EXPORT_SYMBOL_GPL(init_pid_ns);
> >> >
> >> > diff --git a/kernel/pid_namespace.c b/kernel/pid_namespace.c
> >> > index a65ba13..16d6d21 100644
> >> > --- a/kernel/pid_namespace.c
> >> > +++ b/kernel/pid_namespace.c
> >> > @@ -123,6 +123,9 @@ static struct pid_namespace
> >> *create_pid_namespace(struct user_namespace *user_ns
> >> >  	for (i = 1; i < PIDMAP_ENTRIES; i++)
> >> >  		atomic_set(&ns->pidmap[i].nr_free, BITS_PER_PAGE);
> >> >
> >> > +	strncpy(ns->core_pattern, parent_pid_ns->core_pattern,
> >> > +		sizeof(ns->core_pattern));
> >> > +
> >>
> >> This is pretty horrible.  You are giving unprivileged processes the
> >> ability to run an already specified core dump helper in a pid namespace
> >> of their choosing.
> >>
> > Similar problem before patch.
> > In piped core_pattern setting, any panic process will trigger a
> > running of core_dump process.
> 
> That is not at all alike.
> 
> > Comparing to current code, current code maybe more horrible,
> > the guest can destroy host system, and after this patch, the guest
> > can only destroy itself.
> > (As the script in patch description)
> 
> Only if the host is configured to allow itself to be stomped on.  That
> is completely a host configured setting.  Yes the host can configure
> itself in a way that can cause problems.  But that is the hosts problem.
> 
> > Actually it is not so horrible, only the root user can modify code_pattern,
> > and normal user/process have no chance to do bad thing.
> 
> The argument that a bad design is not bad because only root can do X
> does not fly anymore, especially in the presence of containers.
> 
> >> That is not backwards compatible, and it is possible this can lead to
> >> privilege escalation by triciking a privileged dump process to do
> >> something silly because it is running in the wrong pid namespace.
> >>
> > In current code, the dump process is forking from kernel thread,
> > it is in a most-privileged namespace, dumping contents into host's fs,
> > it really cause problem.
> 
> It only dumps into the host's fs if it is configured that way.  When you
> point a gun at your foot and pull the trigger that really causes
> problems as well.  That doesn't make it the gun's problem that it can be
> pointed at feet.
> 
> > Compare to current code, running dump process in container's
> > namespace maybe the right way.
> >
> > The only thing this patch do is letting dump program running in
> > container's namespace instead of host.
> 
> The ability to trick a more privileged program to do the wrong thing is
> most definitely more than only letting the dump program run in the
> container's namespace.  This can be everything up to including getting
> root outside of the container.
> 
> I will admit that if used with a user namespace what you are doing is
> not the worst possible thing to do but it is very definitely a mess.
> 
> >> Similarly the entire concept of forking from the program dumping core
> >> suffers from the same problem but for all other namespaces.
> >>
> >> I was hoping that I would see a justification somewhere in the patch
> >> descriptions describing why this set of decisions could be safe.  I do
> >> not and so I assume this case was not considered.
> >>
> >> If you had managed to fork for the child_reaper of the pid_namespace
> >> that set the core pattern (as has been suggested) there would be some
> >> chance that things would work correctly.
> > Do you mean do fork in kthread(who is running in host's namespace, as
> corrent code)
> > with some special operation to change new thread running in container's
> > namespace?
> >
> >> As you are forking from the program actually dumping core I see no
> >> chance that this patchset is either safe or backwards compatible as
> >> currently written.
> >>
> > Current code have obvious problem, this forking new thread in container's
> > namespace is nothing but safe than host's namespace.
> > At least we need to solve the problem descripted in script in patch
> > description.
> 
> The current code has obvious limitations.  But you can in userspace
> accomplish mush everything you hope to accomplish here, as all of the
> information is available.  It just requires coopeartion.
> 
> > The only thing is backwards compatible, as our discussion in v1 patch,
> > it is the thing we need to change.
> 
> *Laughs*
> 
> This code is so absurd in handling the weird cases that I was hoping
> that someone else would point out how very very bad it is.  The possible
> solutions to the problems have already been discussed to ad nasium and
> you have not used any of those solutions.  Although it does seem the
> kbuild test robot was not too bad in point out how poor your testing of
> these patches was.
> 
> At the end of the day I can break any number of current setups with your
> patches.  Then there are the security implication of confusing
> privileged or somewhat privileged dumping programs.   On the one hand
> your patches are not giving the core dumping program enough privielges
> to write core dumps, on the other hand you are making it possible to
> confuse at least set uid root core dump helpers, leading to privilege
> escalations.
> 
> In no case does what happens during a core dump follow any version of
> the principle of least surprise.
> 
> 
> I will agree that the problem you are trying to solve is a pain point,
> but as I have said before it is a pain point because it is tricky to get
> all of the details right for an real solution.
> 
Let me make a summarize:
You think this way is not acceptable, because the pipe program is running
in the panic-process's namespace context.

And in my view, a pipe program in the host's top level namespace is also
a problem.

Let us think a container, to make it act as a real machine, when a program
panic, linux kernel should dump it into the container's filesystem.

For the kernel, to keep the current way of forking pipe program by kthread,
just let the pipe thread running in the container's namespace, instead the host,
may solve the problem in current kernel.

What is your opinion?

Btw, this patch is trying to solve the problem descripted in thread named:
"piping core dump to a program escapes container" in
http://lists.linuxfoundation.org/pipermail/containers/2015-December/036476.html
Maybe using a userspace tool can make container dump to anywhere,
but for kernel ifself, it is better to solve above problem if we can.

Thanks
Zhaolei