[PATCH RFT 00/16] perf tools: Use generic syscall scripts for all archs

[Charlie Jenkins]

On Mon, Nov 04, 2024 at 02:03:28PM -0800, Ian Rogers wrote:
> On Mon, Nov 4, 2024 at 1:32 PM Charlie Jenkins <charlie at rivosinc.com> wrote:
> >
> > On Mon, Nov 04, 2024 at 10:13:18PM +0100, Arnd Bergmann wrote:
> > > On Mon, Nov 4, 2024, at 22:06, Charlie Jenkins wrote:
> > > > Standardize the generation of syscall headers around syscall tables.
> > > > Previously each architecture independently selected how syscall headers
> > > > would be generated, or would not define a way and fallback onto
> > > > libaudit. Convert all architectures to use a standard syscall header
> > > > generation script and allow each architecture to override the syscall
> > > > table to use if they do not use the generic table.
> > > >
> > > > As a result of these changes, no architecture will require libaudit, and
> > > > so the fallback case of using libaudit is removed by this series.
> > > >
> > > > Testing:
> > > >
> > > > I have tested that the syscall mappings of id to name generation works
> > > > as expected for every architecture, but I have only validated that perf
> > > > trace compiles and runs as expected on riscv, arm64, and x86_64.
> > > >
> > > > Signed-off-by: Charlie Jenkins <charlie at rivosinc.com>
> > >
> > > Thanks for doing this, I had plans to do this myself, but hadn't
> > > completed that bit so far. I'm travelling at the moment, so I'm
> > > not sure I have time to look at it in enough detail this week.
> > >
> > > One problem I ran into doing this previously was the incompatible
> > > format of the tables for x86 and s390, which have conflicting
> > > interpretations of what the '-' character means. It's possible
> > > that this is only really relevant for the in-kernel table,
> > > not the version in tools.
> > >
> >
> > I don't think that is an issue for this usecase because the only
> > information that is taken from the syscall table is the number and the
> > name of the syscall. '-' doesn't appear in either of these columns!
> 
> This is cool stuff. An area that may not be immediately apparent for
> improvement is that the x86-64 build only has access to the 64-bit
> syscall table. Perhaps all the syscall tables should always be built
> and then at runtime the architecture of the perf.data file, etc. used
> to choose the appropriate one. The cleanup to add an ELF host #define
> could help with this:
> https://lore.kernel.org/linux-perf-users/20241017002520.59124-1-irogers@google.com/

Oh that's a great idea! I think these changes will make it more seamless
to make that a reality.

> 
> Ultimately I'd like to see less arch code as it inherently makes cross
> platform worker harder. That doesn't impact this work which I'm happy
> to review.

Yeah I agree. Reducing arch code was the motivation for this change.
There was the issue a couple weeks ago that caused all architectures
that used libaudit to break from commit 7a2fb5619cc1fb53 ("perf trace:
Fix iteration of syscall ids in syscalltbl->entries"), so this change
will eliminate that source of difference between architectures.

- Charlie

> 
> Thanks,
> Ian