Message-ID: <CAP-5=fWrxRvKRxSrYPyiBMZDCfBbb6142ww+er2_mmtkAaGW9A@mail.gmail.com>
Date: Sun, 24 Aug 2025 10:32:16 -0700
From: Ian Rogers <irogers@...gle.com>
To: Sam James <sam@...too.org>
Cc: Ingo Molnar <mingo@...hat.com>, acme@...nel.org, adityag@...ux.ibm.com, 
	adrian.hunter@...el.com, ak@...ux.intel.com, 
	alexander.shishkin@...ux.intel.com, amadio@...too.org, 
	atrajeev@...ux.vnet.ibm.com, bpf@...r.kernel.org, chaitanyas.prakash@....com, 
	changbin.du@...wei.com, charlie@...osinc.com, dvyukov@...gle.com, 
	james.clark@...aro.org, jolsa@...nel.org, justinstitt@...gle.com, 
	kan.liang@...ux.intel.com, kjain@...ux.ibm.com, lihuafei1@...wei.com, 
	linux-kernel@...r.kernel.org, linux-perf-users@...r.kernel.org, 
	llvm@...ts.linux.dev, mark.rutland@....com, mhiramat@...nel.org, 
	morbo@...gle.com, namhyung@...nel.org, nathan@...nel.org, 
	nick.desaulniers+lkml@...il.com, peterz@...radead.org, sesse@...gle.com, 
	song@...nel.org
Subject: Re: [PATCH v5 00/19] Support dynamic opening of capstone/llvm remove BUILD_NONDISTRO

On Sat, Aug 23, 2025 at 3:04 PM Sam James <sam@...too.org> wrote:
>
> Ian Rogers <irogers@...gle.com> writes:
>
> > On Fri, Aug 22, 2025 at 11:52 PM Sam James <sam@...too.org> wrote:
> >>
> >> > A few months ago, objdump was the only way to get
> >> > source line support [0]. Is that still the case?
> >>
> >> ... or is this perhaps handled by "[PATCH v5 18/19] perf srcline:
> >> Fallback between addr2line implementations", in which case, shouldn't
> >> that really land first so people can try the LLVM impl and use the
> >> binutils one if it fails?
> >
> > So my opinion, BUILD_NON_DISTRO isn't supported and the code behind it
> > should go away. Please don't do anything to the contrary or enable it
> > for your distribution - this was supposed to be implied by the name.
>
> We're principally a source-based distribution, so it's not as much of an
> issue.

Agreed, but I don't think the perf tool should be made to be a source
only distribution :-)

> > The forking and running addr2line gets around the license issue that
> > is GPLv3* but comes with a performance overhead. It also has a
> > maintenance overhead supporting llvm and binutil addr2line, when the
> > addr2line output changes things break, etc. (LLVM has been evolving
> > their output but I'm not aware of it breaking things yet). We should
> > (imo) delete the forking and running addr2line support, it fits the
> > billing of something we can do when capstone and libLLVM support
> > aren't there but the code is a hot mess and we don't do exhaustive
> > testing against the many addr2line flavors, the best case is buyer
> > beware. Capstone is derived from libLLVM, I'm not sure it makes sense
> > having 2 libraries for this stuff. There's libLLVM but what it
> > provides through a C API is a mess requiring the C++ shimming. Tbh, I
> > think most of what these libraries provide we should just get over
> > ourselves and provide in perf itself. For example, does it make sense
> > to be trying to add type annotations to objdump output, to just update
> > objdump or have a disassembler library where we can annotate things as
> > we see fit? Library bindings don't break when text output formats get
> > tweaked. Given we're doing so much dwarf processing, do we need a
> > library for that or should that just be in-house? We can side step
> > most of this mess by starting again in python as is being shown in the
> > textual changes that bring with it stuff like console flame graphs:
> > https://lore.kernel.org/lkml/CAP-5=fU=z8kcY4zjezoxSwYf9vczYzHztiMSBvJxdwwBPVWv2Q@mail.gmail.com/
> > So I think long term we make the perf tool minimal with minimal
> > dependencies (ie no addr2line, libLLVM, etc.), we work on having nice
> > stuff in the python stuff where we can reuse or build new libraries
> > for addr2line, objdump-ing, etc. Use >1 thread, use asyncio, etc.
>
> Yeah, this absolutely sounds like the right direction indeed. I'm glad
> to hear it.

Well it takes effort and what drags against that is supporting things
that can't be distributed.

> >
> > For where we are now, ie no python stuff, BUILD_NON_DISTRO should go
> > away as nobody is maintaining it and hasn't for 2 years (what happens
> > when libbfd and libiberty change?)
>
> They don't change often, though. The fixes are usually trivial when they
> do arise.

So just taking the feature tests as an example of the burden, we have:
https://web.git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/build/feature/test-disassembler-four-args.c?h=perf-tools-next
https://web.git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/build/feature/test-disassembler-init-styled.c?h=perf-tools-next
https://web.git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/build/feature/test-libbfd-buildid.c?h=perf-tools-next
https://web.git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/build/feature/test-libbfd.c?h=perf-tools-next
so it isn't just supporting a small libbfd wrapper, it is maintaining
addr2line, objdump, etc. functionality for all these configurations.
As it doesn't get tested, or tested with these features both true and
false, it is liable to bitrot. However, the biggest issue/drag is in
improving/refactoring the tool and having to make sure these new
things work for a situation that in most distributions really, really
doesn't matter.

> > We should focus on making the best
> > of what we have via libraries/tools that are supported - while not
> > forcing the libraries to be there or making the perf binary massive by
> > dragging in say libLLVM. The patch series pushes in that direction and
> > I commend it to the reader.
> >
> > No, reordering the patches to compare performance of binutils doesn't
> > make sense, just build with and without the patch series if you want
> > to do this, but also don't do this as BUILD_NON_DISTRO should go away.
>
> I was asking purely because of the *functionality loss*, though, not
> performance. In the thread I linked from just a few months ago with Ingo
> Molnar, there was a real issue with llvm or capstone-based disassembly
> not showing source information. I'd hit the same problem. Is that fixed now?

Can you link to the issue? Capstone and LLVM don't replace objdump, in
perf there is a list of disassemblers and in your ~/.perfconfig and
you can always set LLVM and capstone to be your least preferred if
they lack a feature you need:
https://web.git.kernel.org/pub/scm/linux/kernel/git/perf/perf-tools-next.git/tree/tools/perf/Documentation/perf-config.txt?h=perf-tools-next#n250

> This is my principal concern rather than the LLVM dependency (even if
> I'd love to avoid that, I understand and appreciate the arguments you're
> making above and intent on future direction).

I think adding a missing feature to the LLVM disassembler is forward
progress, like fixing the BPF disassembly for LLVM done in this
series. I need to know more about the issue but it sounds less of a
libbfd thing and more an objdump -S thing.

This patch series was reposted as I carry it in:
https://github.com/googleprodkernel/linux-perf
to avoid the cost of linking in libLLVM/libcapstone in contexts (data
centers) where that functionality isn't used.

Thanks,
Ian

Powered by blists - more mailing lists