Re: live patching design (was: Re: [PATCH 1/3] sched: add sched_task_call())

[Ingo Molnar <mingo@kernel.org>]

* Josh Poimboeuf <jpoimboe@redhat.com> wrote:

> On Fri, Feb 20, 2015 at 10:46:13PM +0100, Vojtech Pavlik wrote:
> > On Fri, Feb 20, 2015 at 08:49:01PM +0100, Ingo Molnar wrote:
> >
> > > I.e. it's in essence the strong stop-all atomic 
> > > patching model of 'kpatch', combined with the 
> > > reliable avoidance of kernel stacks that 'kgraft' 
> > > uses.
> > 
> > > That should be the starting point, because it's the 
> > > most reliable method.
> > 
> > In the consistency models discussion, this was marked 
> > the "LEAVE_KERNEL+SWITCH_KERNEL" model. It's indeed the 
> > strongest model of all, but also comes at the highest 
> > cost in terms of impact on running tasks. It's so high 
> > (the interruption may be seconds or more) that it was 
> > deemed not worth implementing.
> 
> Yeah, this is way too disruptive to the user.
> 
> Even the comparatively tiny latency caused by kpatch's 
> use of stop_machine() was considered unacceptable by 
> some.

Unreliable, unrobust patching is even more disruptive...

What I think makes it long term fragile is that we combine 
two unrobust, unlikely mechanisms: the chance that a task 
just happens to execute a patched function, with the chance 
that debug information is unreliable.

For example tracing patching got debugged to a fair degree 
because we rely on the patching for actual tracing 
functionality. Even with that relatively robust usage model 
we had our crises ...

I just don't see how a stack backtrace based live patching 
method can become robust in the long run.

> Plus a lot of processes would see EINTR, causing more 
> havoc.

Parking threads safely in user mode does not require the 
propagation of syscall interruption to user-space.

(It does have some other requirements, such as making all 
syscalls interruptible to a 'special' signalling method 
that only live patching triggers - even syscalls that are 
under the normal ABI uninterruptible, such as sys_sync().)

On the other hand, if it's too slow, people will work on 
improving signal propagation latencies: making syscalls 
more readily interruptible and more seemlessly restartable 
has various other advantages beyond live kernel patching.

I.e. it's a win-win scenario and will improve various areas 
of the kernel in terms of syscall interruptability 
latencies.

Thanks,

	Ingo