On Tue, Aug 27, 2019 at 7:26 PM Nicholas Piggin <npig...@gmail.com> wrote:
>
> 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 <npig...@gmail.com>
> >> > wrote:
> >> > > Nick Desaulniers's on August 27, 2019 8:57 am:
> >> > > > On Mon, Aug 26, 2019 at 2:22 PM Nick Desaulniers
> >> > > > <ndesaulni...@google.com> 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
>
+ some Android folks who are looking into libabigail
(root thread for context:
https://lore.kernel.org/lkml/cakwvodnjaapauhtqs7w_vjseybqa0c-tnxrb4xpli0y0soq...@mail.gmail.com/)
I'm only roughly aware of their issue/work, but hopefully they can
collaborate and add their insights to a list of requirements we could
collect, for any kind of replacement to CONFIG_MODVERSIONS/genksyms.
I'm not sure if they were planning on presenting this work at Linux
Plumber's Conference coming up or not, but I hope so. Hopefully that
work will be a general solution for all kernel developers and distros,
not just Android.
This issue is now blocking my compiler upgrade, so I'm going to focus
on just fixing the yacc based parser to understand __attribute__'s on
struct/enum/union declarations.
Speaking with Arnd about this issue, he came up with the test case:
$ echo 'struct __attribute__((packed)) foo { char bar; int baz; };' |
./scripts/genksyms/genksyms -d
Hash table occupancy 0/4096 = 0
$ echo 'struct __attribute__((packed)) foo { char bar; };' |
./scripts/genksyms/genksyms -d
Hash table occupancy 0/4096 = 0
The implication being that because attributes don't parse correctly,
these symbols get a hash value of 0x00 in Module.symvers, meaning we
currently can't detect size changes of packed structs (or any other
attribute-qualified type) and happily proceed loading modules with
different ABIs than the kernel.
I'll probably add some kind of script for unit tests for genksyms,
too. Feels like overkill for something known to be broken where
active replacements are being discussed, but it's blocking my compiler
upgrade for our products, so I'm just going to fix it for now.
--
Thanks,
~Nick Desaulniers
