Re: [PATCH v2 1/2] x86: annotate entry points with type and size

[Roger Pau Monné <roger.pau@xxxxxxxxxx>]

On Tue, May 30, 2023 at 04:23:21PM +0200, Jan Beulich wrote:
> On 30.05.2023 15:21, Roger Pau Monné wrote:
> > On Tue, May 30, 2023 at 10:06:27AM +0200, Jan Beulich wrote:
> >> On 29.05.2023 15:34, Roger Pau Monné wrote:
> >>> On Tue, May 23, 2023 at 01:30:51PM +0200, Jan Beulich wrote:
> >>>> Note that the FB-label in autogen_stubs() cannot be converted just yet:
> >>>> Such labels cannot be used with .type. We could further diverge from
> >>>> Linux'es model and avoid setting STT_NOTYPE explicitly (that's the type
> >>>> labels get by default anyway).
> >>>>
> >>>> Note that we can't use ALIGN() (in place of SYM_ALIGN()) as long as we
> >>>> still have ALIGN.
> >>>
> >>> FWIW, as I'm looking into using the newly added macros in order to add
> >>> annotations suitable for live-patching, I would need to switch some of
> >>> the LABEL usages into it's own functions, as it's not possible to
> >>> livepatch a function that has labels jumped into from code paths
> >>> outside of the function.
> >>
> >> Hmm, I'm not sure what the best way is to overcome that restriction. I'm
> >> not convinced we want to arbitrarily name things "functions".
> > 
> > Any external entry point in the middle of a function-like block will
> > prevent it from being live patched.
> 
> Is there actually any particular reason for this restriction? As long
> as old and new code has the same external entry points, redirecting
> all old ones to their new counterparts would seem feasible.

Yes, that was another option, we could force asm patching to always be
done with a jump (instead of in-place) and then add jumps at the old
entry point addresses in order to redirect to the new addresses.

Or assert that the addresses of any symbols inside the function is not
changed in order to do in-place replacement of code.

> > If you want I can try to do a pass on top of your patch and see how
> > that would end up looking.  I'm attempting to think about other
> > solutions, but every other solution seems quite horrible.
> 
> Right, but splitting functions into piecemeal fragments isn't going
> to be very nice either.

I'm not sure how much splitting would be required TBH.

> >>>> +
> >>>> +#define FUNC(name, algn...) \
> >>>> +        SYM(name, FUNC, GLOBAL, LAST(16, ## algn), 0x90)
> >>>
> >>> A rant, should the alignment of functions use a different padding?
> >>> (ie: ret or ud2?) In order to prevent stray jumps falling in the
> >>> padding and fall trough into the next function.  That would also
> >>> prevent the implicit fall trough used in some places.
> >>
> >> Yes, but that's a separate topic (for which iirc patches are pending
> >> as well, just of course not integrated with the work here. There's
> >> the slight risk of overlooking some "fall-through" case ...
> > 
> > Oh, OK, wasn't aware patches are floating for this already, just came
> > across it while reviewing.
> 
> Well, those don't cover padding yet, but they deal with straight-line
> speculation past RET or JMP.

Introducing the helpers does make it easy to convert the padding for
all the existing users at least.

> >>>>          sti
> >>>>          call  do_softirq
> >>>>          jmp   compat_test_all_events
> >>>>  
> >>>> -        ALIGN
> >>>>  /* %rbx: struct vcpu, %rdx: struct trap_bounce */
> >>>> -.Lcompat_process_trapbounce:
> >>>> +LABEL_LOCAL(.Lcompat_process_trapbounce)
> >>>
> >>> It's my understanding that here the '.L' prefix is pointless, since
> >>> LABEL_LOCAL() will forcefully create a symbol for the label due to the
> >>> usage of .type?
> >>
> >> I don't think .type has this effect. There's certainly no such label in
> >> the symbol table of the object file I have as a result.
> > 
> > I was expecting .type to force the creation of a symbol, so the '.L'
> > prefix does prevent the symbol from being created even if .type is
> > specified.
> > 
> > Shouldn't the assembler complain that we are attempting to set a type
> > for a not present symbol?
> 
> But .L symbols are still normal symbols to gas, just that it knows to not
> emit them to the symbol table (unless there's a need, e.g. through a use
> in a relocation that cannot be expressed as section-relative one). It
> could flag the pointless use, but then it may get this wrong if in the
> end the symbol does need emitting.

Thanks for the explanation.

Roger.