On Mon, Jun 30, 2025 at 09:31:44AM +0200, Peter Zijlstra wrote:
> On Fri, Jun 27, 2025 at 10:34:15AM -0700, Josh Poimboeuf wrote:
> > On Fri, Jun 27, 2025 at 12:53:28PM +0200, Peter Zijlstra wrote:
> > > On Thu, Jun 26, 2025 at 04:55:29PM -0700, Josh Poimboeuf wrote:
> > > > An upcoming patch will add the SHF_MERGE flag to x86 __jump_table and
> > > > __bug_table so their entry sizes can be defined in inline asm.
> > > > 
> > > > However, those sections have SHF_WRITE, which the Clang linker (lld)
> > > > explicitly forbids combining with SHF_MERGE.
> > > > 
> > > > Those sections are modified at runtime and must remain writable.  While
> > > > SHF_WRITE is ignored by vmlinux, it's still needed for modules.
> > > > 
> > > > To work around the linker interference, remove SHF_WRITE during
> > > > compilation and restore it after linking the module.
> > > 
> > > This is vile... but I'm not sure I have a better solution.
> > > 
> > > Eventually we should get the toolchains fixed, but we can't very well
> > > mandate clang-21+ to build x86 just yet.
> > 
> > Yeah, I really hate this too.  I really tried to find something better,
> > including mucking with the linker script, but this was unfortunately the
> > only thing that worked.
> > 
> > Though, looking at it again, I realize we can localize the pain to Clang
> > (and the makefile) by leaving the code untouched and instead strip
> > SHF_WRITE before the link and re-add it afterwards.  Then we can tie
> > this horrible hack to specific Clang versions when it gets fixed.
> 
> Oh yeah, that might be nicer indeed!

I ended up giving up on this approach.  It was just too painful trying
to get the toolchain to cooperate.  It does kind of make sense that
SHF_WRITE and SHF_MERGE are incompatible.

What I really needed was a way to set entsize from assembly, but that
doesn't seem possible without SHF_MERGE.

For now I'm just going with marking the beginning of each special
section entry with an objtool annotation.  Then objtool can deduce the
sizes from the distance between the annotations.

-- 
Josh

Reply via email to