Re: [PATCH 2/2] kernel: add support for live patching

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Fri, Nov 07, 2014 at 07:21:03PM +0100, Petr Mladek wrote:
> On Thu 2014-11-06 10:57:48, Seth Jennings wrote:
> > On Thu, Nov 06, 2014 at 04:51:02PM +0100, Jiri Slaby wrote:
> > > On 11/06/2014, 03:39 PM, Seth Jennings wrote:
> > > > +/*************************************
> > > > + * Core structures
> > > > + ************************************/
> > > > +
> > > > +/*
> > > > + * lp_ structs vs lpc_ structs
> > > > + *
> > > > + * For each element (patch, object, func) in the live-patching code,
> > > > + * there are two types with two different prefixes: lp_ and lpc_.
> > > > + *
> > > > + * Structures used by the live-patch modules to register with this core module
> > > > + * are prefixed with lp_ (live patching).  These structures are part of the
> > > > + * registration API and are defined in livepatch.h.  The structures used
> > > > + * internally by this core module are prefixed with lpc_ (live patching core).
> > > > + */
> > > 
> > > I am not sure if the separation and the allocations/kobj handling are
> > > worth it. It makes the code really less understandable. Can we have just
> > > struct lip_function (don't unnecessarily abbreviate), lip_objectfile
> > > (object is too generic, like Java object) and lip_patch containing all
> > > the needed information? It would clean up the code a lot. (Yes, we would
> > > have profited from c++ here.)
> > 
> > I looked at doing this and this is actually what we did in kpatch.  We
> > made one structure that had "private" members that the user wasn't
> > suppose to access that were only used in the core.  This was messy
> > though.  Every time you wanted to add a "private" field to the struct so
> > the core could do something new, you were changing the API to the patch
> > modules as well.  While copying the data into an internal structure does
> > add code and opportunity for errors, that functionality is localized
> > into functions that are specifically tasked with taking care of that.
> > So the risk is minimized and we gain flexibility within the core and
> > more self-documenting API structures.
> 
> I am not sure if the modified API is really such a big limit. The
> modules initialize the needed members using ".member = value".
> Also we do not need to take care of API/ABI backward compatibility because
> there is very strict dependency between patches and the patched
> kernel.

Our patch module generation tool (kpatch-build) relies on the API as
well, so we should try to keep the API as stable as possible.  At least
until we can put kpatch-build (or something like it) into the kernel
tree.

-- 
Josh