lkml.org 
[lkml]   [2016]   [May]   [4]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC PATCH v2 17/18] livepatch: change to a per-task consistency model
On Wed, May 04, 2016 at 10:42:23AM +0200, Petr Mladek wrote:
> On Thu 2016-04-28 15:44:48, Josh Poimboeuf wrote:
> > Change livepatch to use a basic per-task consistency model. This is the
> > foundation which will eventually enable us to patch those ~10% of
> > security patches which change function or data semantics. This is the
> > biggest remaining piece needed to make livepatch more generally useful.
>
> > --- a/kernel/fork.c
> > +++ b/kernel/fork.c
> > @@ -76,6 +76,7 @@
> > #include <linux/compiler.h>
> > #include <linux/sysctl.h>
> > #include <linux/kcov.h>
> > +#include <linux/livepatch.h>
> >
> > #include <asm/pgtable.h>
> > #include <asm/pgalloc.h>
> > @@ -1586,6 +1587,8 @@ static struct task_struct *copy_process(unsigned long clone_flags,
> > p->parent_exec_id = current->self_exec_id;
> > }
> >
> > + klp_copy_process(p);
>
> I am in doubts here. We copy the state from the parent here. It means
> that the new process might still need to be converted. But at the same
> point print_context_stack_reliable() returns zero without printing
> any stack trace when TIF_FORK flag is set. It means that a freshly
> forked task might get be converted immediately. I seems that boot
> operations are always done when copy_process() is called. But
> they are contradicting each other.
>
> I guess that print_context_stack_reliable() should either return
> -EINVAL when TIF_FORK is set. Or it should try to print the
> stack of the newly forked task.
>
> Or do I miss something, please?

Ok, I admit it's confusing.

A newly forked task doesn't *have* a stack (other than the pt_regs frame
it needs for the return to user space), which is why
print_context_stack_reliable() returns success with an empty array of
addresses.

For a little background, see the second switch_to() macro in
arch/x86/include/asm/switch_to.h. When a newly forked task runs for the
first time, it returns from __switch_to() with no stack. It then jumps
straight to ret_from_fork in entry_64.S, calls a few C functions, and
eventually returns to user space. So, assuming we aren't patching entry
code or the switch_to() macro in __schedule(), it should be safe to
patch the task before it does all that.

With the current code, if an unpatched task gets forked, the child will
also be unpatched. In theory, we could go ahead and patch the child
then. In fact, that's what I did in v1.9.

But in v1.9 discussions it was pointed out that someday maybe the
ret_from_fork stuff will get cleaned up and instead the child stack will
be copied from the parent. In that case the child should inherit its
parent's patched state. So we decided to make it more future-proof by
having the child inherit the parent's patched state.

So, having said all that, I'm really not sure what the best approach is
for print_context_stack_reliable(). Right now I'm thinking I'll change
it back to return -EINVAL for a newly forked task, so it'll be more
future-proof: better to have a false positive than a false negative.
Either way it will probably need to be changed again if the
ret_from_fork code gets cleaned up.

> > +
> > spin_lock(&current->sighand->siglock);
> >
> > /*
>
> [...]
>
> > diff --git a/kernel/livepatch/transition.c b/kernel/livepatch/transition.c
> > new file mode 100644
> > index 0000000..92819bb
> > --- /dev/null
> > +++ b/kernel/livepatch/transition.c
> > +/*
> > + * This function can be called in the middle of an existing transition to
> > + * reverse the direction of the target patch state. This can be done to
> > + * effectively cancel an existing enable or disable operation if there are any
> > + * tasks which are stuck in the initial patch state.
> > + */
> > +void klp_reverse_transition(void)
> > +{
> > + struct klp_patch *patch = klp_transition_patch;
> > +
> > + klp_target_state = !klp_target_state;
> > +
> > + /*
> > + * Ensure that if another CPU goes through the syscall barrier, sees
> > + * the TIF_PATCH_PENDING writes in klp_start_transition(), and calls
> > + * klp_patch_task(), it also sees the above write to the target state.
> > + * Otherwise it can put the task in the wrong universe.
> > + */
> > + smp_wmb();
> > +
> > + klp_start_transition();
> > + klp_try_complete_transition();
>
> It is a bit strange that we keep the work scheduled. It might be
> better to use
>
> mod_delayed_work(system_wq, &klp_work, 0);

True, I think that would be better.

> Which triggers more ideas from the nitpicking deparment:
>
> I would move the work definition from core.c to transition.c because
> it is closely related to klp_try_complete_transition();

That could be good, but there's a slight problem: klp_work_fn() requires
klp_mutex, which is static to core.c. It's kind of nice to keep the use
of the mutex in core.c only.

> When on it. I would make it more clear that the work is related
> to transition.

How would you recommend doing that? How about:

- rename "klp_work" -> "klp_transition_work"
- rename "klp_work_fn" -> "klp_transition_work_fn"

?

> Also I would call queue_delayed_work() directly
> instead of adding the klp_schedule_work() wrapper. The delay
> might be defined using a constant, e.g.
>
> #define KLP_TRANSITION_DELAY round_jiffies_relative(HZ)
>
> queue_delayed_work(system_wq, &klp_transition_work, KLP_TRANSITION_DELAY);

Sure.

> Finally, the following is always called right after
> klp_start_transition(), so I would call it from there.
>
> if (!klp_try_complete_transition())
> klp_schedule_work();

Except for when it's called by klp_reverse_transition(). And it really
depends on whether we want to allow transition.c to use the mutex. I
don't have a strong opinion either way, I may need to think about it
some more.

> > +
> > + patch->enabled = !patch->enabled;
> > +}
> > +
>
> It is really great work! I am checking this patch from left, right, top,
> and even bottom and all seems to work well together.

Great! Thanks a lot for the thorough review!

--
Josh

\
 
 \ /
  Last update: 2016-05-04 18:01    [W:1.154 / U:0.980 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site