Re: a bug in genksysms/CONFIG_MODVERSIONS w/ __attribute__((foo))?

[Nicholas Piggin]

Ben Hutchings's on August 28, 2019 1:34 am:
> On Tue, 2019-08-27 at 22:42 +1000, Nicholas Piggin wrote:
>> Masahiro Yamada's on August 27, 2019 8:49 pm:
>> > Hi.
>> > 
>> > On Tue, Aug 27, 2019 at 6:59 PM Nicholas Piggin <npiggin@xxxxxxxxx> wrote:
>> > > Nick Desaulniers's on August 27, 2019 8:57 am:
>> > > > On Mon, Aug 26, 2019 at 2:22 PM Nick Desaulniers
>> > > > <ndesaulniers@xxxxxxxxxx> wrote:
>> > > > > I'm looking into a linkage failure for one of our device kernels, and
>> > > > > it seems that genksyms isn't producing a hash value correctly for
>> > > > > aggregate definitions that contain __attribute__s like
>> > > > > __attribute__((packed)).
>> > > > > 
>> > > > > Example:
>> > > > > $ echo 'struct foo { int bar; };' | ./scripts/genksyms/genksyms -d
>> > > > > Defn for struct foo == <struct foo { int bar ; } >
>> > > > > Hash table occupancy 1/4096 = 0.000244141
>> > > > > $ echo 'struct __attribute__((packed)) foo { int bar; };' |
>> > > > > ./scripts/genksyms/genksyms -d
>> > > > > Hash table occupancy 0/4096 = 0
>> > > > > 
>> > > > > I assume the __attribute__ part isn't being parsed correctly (looks
>> > > > > like genksyms is a lex/yacc based C parser).
>> > > > > 
>> > > > > The issue we have in our out of tree driver (*sadface*) is basically a
>> > > > > EXPORT_SYMBOL'd function whose signature contains a packed struct.
>> > > > > 
>> > > > > Theoretically, there should be nothing wrong with exporting a function
>> > > > > that requires packed structs, and this is just a bug in the lex/yacc
>> > > > > based parser, right?  I assume that not having CONFIG_MODVERSIONS
>> > > > > coverage of packed structs in particular could lead to potentially
>> > > > > not-fun bugs?  Or is using packed structs in exported function symbols
>> > > > > with CONFIG_MODVERSIONS forbidden in some documentation somewhere I
>> > > > > missed?
>> > > > 
>> > > > Ah, looks like I'm late to the party:
>> > > > https://lwn.net/Articles/707520/
>> > > 
>> > > Yeah, would be nice to do something about this.
>> > 
>> > modversions is ugly, so it would be great if we could dump it.
>> > 
>> > > IIRC (without re-reading it all), in theory distros would be okay
>> > > without modversions if they could just provide their own explicit
>> > > versioning. They take care about ABIs, so they can version things
>> > > carefully if they had to change.
> 
> Debian doesn't currently have any other way of detecting ABI changes
> (other than eyeballing diffs).
> 
> I know there have been proposals of using libabigail for this instead,
> but I'm not sure how far those progressed.
> 
>> > We have not provided any alternative solution for this, haven't we?
>> > 
>> > In your patch (https://lwn.net/Articles/707729/),
>> > you proposed CONFIG_MODULE_ABI_EXPLICIT.
>> 
>> Right, that was just my first proposal, but I am not confident that I
>> understood everybody's requirements. I don't think the distro people
>> had much time to to test things out.
>> 
>> One possible shortcoming with that patch is no per-symbol version. The 
>> distro may break an ABI for a security fix, but they don't want to break
>> all out of tree modules if it's an obscure ABI.
> 
> Right, for example the KVM kABI is only meant for in-tree modules (like
> kvm_intel) and in Debian we do not change the "ABI version" and require
> rebuilding out-of-tree modules just because that ABI changes. 
> Currently we maintain explicit lists of exported symbols and exporting
> modules for which we ignore ABI changes at build time.
> 
>> The counter argument to 
>> that is they should just rename the symbol in their kernel for such 
>> cases, so I didn't implement it without somebody describing a good
>> requirement.
> [...]
> 
> Sometimes it is just a single function that changes, but often a
> structure change can affect large numbers of functions.  For example,
> if KVM adds a member to an operations struct that can indirectly change
> the ABI for most of its exported functions.  We wouldn't want to change
> the ABI version but would still want to prevent loading mismatched kvm
> and kvm_intel versions.  It would be a lot more work to change all of
> the affected function names.

You could change just a single symbol name though :)

> An alternative to symbol version matching that I think would work for
> us is: if a module's exports or imports match the "changes ignored"
> list then the module can only be loaded on the exact version of the
> kernel, otherwise it only needs to match the ABI version.  I think that
> would avoid the need for carrying symbol versions, but we would still
> need a build-time ABI check and a way of flagging which symbols need
> the tighter version match.

Just trying to think how best to express that.

[ Aside, the whole symbol name resolution linking stuff does matching on 
  on any number of ~arbitrary strings that you can generate as you like, 
  and symbol tables are something that all existing tools and libs 
  understand.

  So I strongly favour using that as the back end for our "version"
  resolution system _if at all possible_ rather than adding these extra
  bits of crud that really just do the same thing. At least for a first
  pass, I don't want to over-engineer things.

  Then it hopefully becomes a matter of adding some helper macros and
  build facilities on top of that which can contain everyone's
  requirements mostly within .config and perhaps a very small patch.
  A bit more work with preprocessor macros etc is far preferable to
  linking and loading "features" IMO]

Back to your case, is it sufficient to have just an internal and an
external module version where the kernel provides both and your in-tree
modules match on the internal, others match on external?

Thanks,
Nick