[PATCH] modpost: support arbitrary symbol length in modversion

[Andrea Righi]

On Thu, Mar 16, 2023 at 01:18:23AM +0900, Masahiro Yamada wrote:
> On Tue, Mar 14, 2023 at 11:38 PM Andrea Righi
> <andrea.righi at canonical.com> wrote:
> >
> > On Mon, Mar 13, 2023 at 11:09:31PM +0100, Andrea Righi wrote:
> > > On Mon, Mar 13, 2023 at 11:02:34PM +0100, Michal Suchánek wrote:
> > > > On Mon, Mar 13, 2023 at 10:53:34PM +0100, Andrea Righi wrote:
> > > > > On Mon, Mar 13, 2023 at 10:48:53PM +0100, Michal Suchánek wrote:
> > > > > > Hello,
> > > > > >
> > > > > > On Mon, Mar 13, 2023 at 09:32:16PM +0100, Andrea Righi wrote:
> > > > > > > On Wed, Jan 11, 2023 at 04:11:51PM +0000, Gary Guo wrote:
> > > > > > > > Currently modversion uses a fixed size array of size (64 - sizeof(long))
> > > > > > > > to store symbol names, thus placing a hard limit on length of symbols.
> > > > > > > > Rust symbols (which encodes crate and module names) can be quite a bit
> > > > > > > > longer. The length limit in kallsyms is increased to 512 for this reason.
> > > > > > > >
> > > > > > > > It's a waste of space to simply expand the fixed array size to 512 in
> > > > > > > > modversion info entries. I therefore make it variably sized, with offset
> > > > > > > > to the next entry indicated by the initial "next" field.
> > > > > > > >
> > > > > > > > In addition to supporting longer-than-56/60 byte symbols, this patch also
> > > > > > > > reduce the size for short symbols by getting rid of excessive 0 paddings.
> > > > > > > > There are still some zero paddings to ensure "next" and "crc" fields are
> > > > > > > > properly aligned.
> > > > > > > >
> > > > > > > > This patch does have a tiny drawback that it makes ".mod.c" files generated
> > > > > > > > a bit less easy to read, as code like
> > > > > > > >
> > > > > > > >     "\x08\x00\x00\x00\x78\x56\x34\x12"
> > > > > > > >     "symbol\0\0"
> > > > > > > >
> > > > > > > > is generated as opposed to
> > > > > > > >
> > > > > > > >     { 0x12345678, "symbol" },
> > > > > > > >
> > > > > > > > because the structure is now variable-length. But hopefully nobody reads
> > > > > > > > the generated file :)
> > > > > > > >
> > > > > > > > Link: b8a94bfb3395 ("kallsyms: increase maximum kernel symbol length to 512")
> > > > > > > > Link: https://github.com/Rust-for-Linux/linux/pull/379
> > > > > > > >
> > > > > > > > Signed-off-by: Gary Guo <gary at garyguo.net>
> > > > > > >
> > > > > > > Is there any newer version of this patch?
> > > > > > >
> > > > > > > I'm doing some tests with it, but I'm getting boot failures on ppc64
> > > > > > > with this applied (at boot kernel is spitting out lots of oops'es and
> > > > > > > unfortunately it's really hard to copy paste or just read them from the
> > > > > > > console).
> > > > > >
> > > > > > Are you using the ELF ABI v1 or v2?
> > > > > >
> > > > > > v1 may have some additional issues when it comes to these symbol tables.
> > > > > >
> > > > > > Thanks
> > > > > >
> > > > > > Michal
> > > > >
> > > > > I have CONFIG_PPC64_ELF_ABI_V2=y in my .config, so I guess I'm using v2.
> > > > >
> > > > > BTW, the issue seems to be in dedotify_versions(), as a silly test I
> > > > > tried to comment out this function completely to be a no-op and now my
> > > > > system boots fine (but I guess I'm probably breaking something else).
> > > >
> > > > Probably not. You should not have the extra leading dot on ABI v2. So if
> > > > dedotify does something that means something generates and then expects
> > > > back symbols with a leading dot, and this workaround for ABI v1 breaks
> > > > that. Or maybe it is called when it shouldn't.
> > >
> > > Hm.. I'll add some debugging to this function to see what happens exactly.
> >
> > Alright I've done more tests across different architectures. My problem
> > with ppc64 is that this architecture is evaluating sechdrs[i].sh_size
> > using get_stubs_size(), that apparently can add some extra padding, so
> > doing (vers + vers->next < end) isn't a reliable check to determine the
> > end of the variable array, because sometimes "end" can be greater than
> > the last "vers + vers->next" entry.
> 
> 
> I am not familiar enough with ppc, so I may be misundering.
> 
> Checking the for-loop in module_frob_arch_sections(),
> they seem to be orthogonal to me.
> 
> dedotify_versions() is only called for the "__versions" section.
> get_stubs_size() only affects the ".stubs" section.
> I did not get how they are related to each other.
> 
> 
> BTW, we decided to not go in this direction in the former discussion.
> I am not sure how much effort is needed to track down the issue
> in this version.
> 
> If we add new sections to keep the backward compatibility
> for the current "__versions", this issue may not exist.

I think need to investigate more on this. But according to the results
(and some extra printk()'s that I added) it seems that sechdrs[].sh_size
is bigger than expected, so using that to determine the end of the array
is not reliable and it systematically breaks the boot on the ppc64
system that I'm currently using.

Anyway, I know that we decided to drop this approach, I'm actually using
this patch to see if we can enable CONFIG_MODVERSIONS with CONFIG_RUST
enabled, that's why I'm exploring the possibility to have a variable
length for module symbols.

Hopefully we'll come up with a proper approach/solution, because it'd be
really useful to have MODVERSIONS and RUST both enabled.

Thanks,
-Andrea