Re: [PATCH v2 dwarves 0/8] Switch BTF loading and encoding to libbpf APIs

From: Andrii Nakryiko <andrii.nakryiko@gmail.com>
To: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Andrii Nakryiko <andrii@kernel.org>, Hao Luo <haoluo@google.com>,
	Oleg Rombakh <olegrom@google.com>,
	dwarves@vger.kernel.org, bpf <bpf@vger.kernel.org>,
	Kernel Team <kernel-team@fb.com>,
	Alexei Starovoitov <ast@kernel.org>,
	Tony Ambardar <tony.ambardar@gmail.com>,
	Ilya Leoshkevich <iii@linux.ibm.com>,
	David Marcinkovic <david.marcinkovic@sartura.hr>,
	Luka Perkov <luka.perkov@sartura.hr>,
	Borna Cafuk <borna.cafuk@sartura.hr>,
	Juraj Vijtiuk <juraj.vijtiuk@sartura.hr>
Subject: Re: [PATCH v2 dwarves 0/8] Switch BTF loading and encoding to libbpf APIs
Date: Fri, 9 Oct 2020 10:59:20 -0700	[thread overview]
Message-ID: <CAEf4BzZa7WGDpBxHKNvO+SWv2=J2mGyQZuuMTkGWkRWxz+KLgg@mail.gmail.com> (raw)
In-Reply-To: <20201009162243.GD322246@kernel.org>

On Fri, Oct 9, 2020 at 9:22 AM Arnaldo Carvalho de Melo <acme@kernel.org> wrote:
>
> Em Thu, Oct 08, 2020 at 04:39:52PM -0700, Andrii Nakryiko escreveu:
> > This patch set switches pahole to use libbpf-provided BTF loading and encoding
> > APIs. This reduces pahole's own BTF encoding code, speeds up the process,
> > reduces amount of RAM needed for DWARF-to-BTF conversion. Also, pahole finally
> > gets support to generating BTF for cross-compiled ELF binaries with different
> > endianness (patch #8).
> >
> > Additionally, patch #3 fixes previously missed problem with invalid array
> > index type generation.
> >
> > Patches #4-7 are speeding up DWARF-to-BTF convertion/dedup pretty
> > significantly, saving overall about 9 seconds out of current 27 or so.
> >
> > Patch #5 revamps how per-CPU BTF variables are emitted, eliminating repeated
> > and expensive looping over ELF symbols table. The critical detail that took
> > few hours of investigation is that when DW_AT_variable has
> > DW_AT_specification, variable address (to correlate with symbol's address) has
> > to be taken before specification is followed.
> >
> > More details could be found in respective patches.
> >
> > v1->v2:
> >   - rebase on latest dwarves master and fix var->spec's address problem.
>
> Thanks, I applied all of them, tested and reproduced the performance
> gains, great work!

Great, thanks a lot, Arnaldo!

Next step is adding BTF to kernel modules, where module's BTF will be
an "extension" of vmlinux's BTF, with only a minimal set of new types
used/added in the module, that are not available in vmlinux. This
should make per-module BTF really tiny.

>
> I'll do some more testing on encoding a vmlinux for some big endian arch
> on my x86_64 workstation and then push things publicly.
>
> If Hao find any issues we can fix in a follow up patch.
>
> I also added the people involved in the discussion about cross builds
> failing, please take a look, I'm pushing now to a tmp.libbtf_encoder so
> that you can test it from there, ok?
>
> - Arnaldo
>
> > Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
> >
> > Andrii Nakryiko (8):
> >   btf_loader: use libbpf to load BTF
> >   btf_encoder: use libbpf APIs to encode BTF type info
> >   btf_encoder: fix emitting __ARRAY_SIZE_TYPE__ as index range type
> >   btf_encoder: discard CUs after BTF encoding
> >   btf_encoder: revamp how per-CPU variables are encoded
> >   dwarf_loader: increase the size of lookup hash map
> >   strings: use BTF's string APIs for strings management
> >   btf_encoder: support cross-compiled ELF binaries with different
> >     endianness
> >
> >  btf_encoder.c  | 370 +++++++++++++++------------
> >  btf_loader.c   | 244 +++++++-----------
> >  ctf_encoder.c  |   2 +-
> >  dwarf_loader.c |   2 +-
> >  libbtf.c       | 661 +++++++++++++++++++++----------------------------
> >  libbtf.h       |  41 ++-
> >  libctf.c       |  14 +-
> >  libctf.h       |   4 +-
> >  pahole.c       |   2 +-
> >  strings.c      |  91 +++----
> >  strings.h      |  32 +--
> >  11 files changed, 645 insertions(+), 818 deletions(-)
> >
> > --
> > 2.24.1
> >
>
> --
>
> - Arnaldo