Re: [PATCHv4 2/3] kernel: add support for live patching

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Mon, Dec 01, 2014 at 02:31:35PM +0100, Miroslav Benes wrote:
> On Wed, 26 Nov 2014, Josh Poimboeuf wrote:
> 
> > Hi Miroslav,
> > 
> > Just addressing one of your comments below.  I'll let Seth respond to
> > the others :-)
> > 
> > On Wed, Nov 26, 2014 at 03:19:17PM +0100, Miroslav Benes wrote:
> > > > +/**
> > > > + * struct klp_func - function structure for live patching
> > > > + * @old_name:	name of the function to be patched
> > > > + * @new_func:	pointer to the patched function code
> > > > + * @old_addr:	a hint conveying at what address the old function
> > > > + *		can be found (optional, vmlinux patches only)
> > > > + */
> > > > +struct klp_func {
> > > > +	/* external */
> > > > +	const char *old_name;
> > > > +	void *new_func;
> > > > +	/*
> > > > +	 * The old_addr field is optional and can be used to resolve
> > > > +	 * duplicate symbol names in the vmlinux object.  If this
> > > > +	 * information is not present, the symbol is located by name
> > > > +	 * with kallsyms. If the name is not unique and old_addr is
> > > > +	 * not provided, the patch application fails as there is no
> > > > +	 * way to resolve the ambiguity.
> > > > +	 */
> > > > +	unsigned long old_addr;
> > > 
> > > I wonder if we really need old_addr as an external field. I assume that 
> > > userspace tool in kpatch use it as a "hint" for kernel part and thus 
> > > kallsyms is not needed there (and it solves ambiguity problem as well). 
> > > But I am not sure if it is gonna be the same in upstream. When kernel is 
> > > randomized (CONFIG_RANDOMIZE_BASE is set to 'y', though default is 'n') 
> > > old_addr is not usable (and we throw it away in the code). Without 
> > > old_addr being set by the user we could spare some of code (calls to 
> > > klp_verify_vmlinux_symbol and such). 
> > 
> > Even with CONFIG_RANDOMIZE_BASE, the function offsets will be the same
> > regardless of the base address.  So we could still use old_addr to
> > determine the offset.
> > 
> > > So the question is whether future userspace tool in upstream would need it 
> > > and would use it. Please note that I do not mean it as a kpatch or kgraft 
> > > way to do things, I'm just not sure about old_addr being "public" and want 
> > > do discuss the options.
> > > 
> > > The ambiguity of symbols was discussed in some other thread in lkml in 
> > > october (I guess) with no conclusion IIRC...
> > 
> > We need to resolve ambiguity somehow, and old_addr is a way to do that.
> > Do you have any other ideas?
> 
> Unfortunately I don't.
> 
> But similarly we don't deal with ambiguity in modules either. And it is 
> (at least theoretically) possible. Two static functions of the same name 
> in two different .c files which the final module is linked from. You have 
> to use kallsyms and it would get confused. Maybe this sounds odd but it 
> could happen. 

True, this is a remote possibility, but we haven't run into this issue
yet.  If it becomes a problem, we can try to come up with another way to
resolve duplicates.

Here's one idea: since the symbols are always listed in the same order
in kallsyms (per-object), instead of old_addr we could have sym_idx.  A
sym_idx of 2 could mean "I want the 2nd occurence of foo in the object's
kallsyms list".

However I'd rather keep our current old_addr scheme for now, since it's
what we have implemented already.  And there are plenty of more
important things we need to do first.

> Thus the old_addr value is not general protection (as modules are still 
> affected) and it is questionable whether the user should use it.

It's not really protection, since even if you don't specify old_addr and
you rely on kallsyms, klp_find_symbol will return an error if there are
any duplicates.

It's really just a way to increase the size of the set of functions
which can be patched (duplicately named functions).

We also rely on something similar for relocations: klp_reloc.src.  It's
even more important there, since duplicately named static objects are
more common than duplicately named functions.

> I do not have strong opinion on this and if no one else shares my 
> thoughts, I am not against.
> 
> Mira

-- 
Josh