This is a v2 of the patchset aiming to speed up parallel BTF encoding
when reproducible_build flag is set (see link [1]).
In comparison to v1:
* patch #2 adding section-relative addresses to elf_functions is
removed as unrelated [2]
* patch #9 [3] is replaced with patches #8, #9 and #10 (the biggest
and most important in this series)
Patch #10 rewrites multithreading implementation to job/worker
model. See the details in the commit message.
The ./tests/tests pass with a vmlinux build on bpf-next.
I also confrimed that the reproducible bpftool dump of BTF produced
for vmlinux is identical between this patch series and pahole/next.
With this patch series, the performance of parallel BTF encoding is
comparable to non-reproducible runs on pahole/next. Depending on the
number of threads and allowed memory usage (indirectly controlled by
max_decoded_cus parameter of the queue in the dwarf_loader.c), it may
be a little slower or a little faster.
Note that the number of CPU cycles is significantly less, although the
wall-clock time is somewhat greater for -j24, as reported by perf.
See sample measurements below (host nproc=24).
This patch (always reproducible)
-j1 mem 842020 Kb, time 6.31 sec
-j3 mem 864604 Kb, time 2.90 sec
-j6 mem 927760 Kb, time 2.21 sec
-j12 mem 1026616 Kb, time 2.29 sec
-j24 mem 1188448 Kb, time 2.36 sec
-j48 mem 1462656 Kb, time 2.48 sec
Performance counter stats for '/home/theihor/dev/dwarves/build/pahole -J -j24 --btf_features=encode_force,var,float,enum64,decl_tag,type_tag,optimized_func,consistent_func,decl_tag_kfuncs,reproducible_build --btf_encode_detached=/dev/null --lang_exclude=rust /home/theihor/git/kernel.org/bpf-next/kbuild-output/.tmp_vmlinux1' (13 runs):
46,771,092,586 cycles:u ( +- 0.17% )
2.36785 +- 0.00503 seconds time elapsed ( +- 0.21% )
pahole/next (1cb4202) non-reproducible
-j1 mem 834004 Kb, time 6.25 sec
-j3 mem 976480 Kb, time 3.21 sec
-j6 mem 1081432 Kb, time 2.36 sec
-j12 mem 1161252 Kb, time 2.07 sec
-j24 mem 1303060 Kb, time 2.13 sec
-j48 mem 1537800 Kb, time 2.39 sec
Performance counter stats for '/home/theihor/dev/dwarves/build/pahole -J -j24 --btf_features=encode_force,var,float,enum64,decl_tag,type_tag,optimized_func,consistent_func,decl_tag_kfuncs --btf_encode_detached=/dev/null --lang_exclude=rust /home/theihor/git/kernel.org/bpf-next/kbuild-output/.tmp_vmlinux1' (13 runs):
60,436,382,442 cycles:u ( +- 0.22% )
2.2024 +- 0.0151 seconds time elapsed ( +- 0.68% )
pahole/next (1cb4202) reproducible
-j1 mem 4745764 Kb, time 7.64 sec
-j3 mem 4744556 Kb, time 3.95 sec
-j6 mem 4744592 Kb, time 2.98 sec
-j12 mem 4744680 Kb, time 2.99 sec
-j24 mem 4745252 Kb, time 2.99 sec
-j48 mem 4744520 Kb, time 2.98 sec
Performance counter stats for '/home/theihor/dev/dwarves/build/pahole -J -j24 --btf_features=encode_force,var,float,enum64,decl_tag,type_tag,optimized_func,consistent_func,decl_tag_kfuncs,reproducible_build --btf_encode_detached=/dev/null --lang_exclude=rust /home/theihor/git/kernel.org/bpf-next/kbuild-output/.tmp_vmlinux1' (13 runs):
38,155,725,721 cycles:u ( +- 0.29% )
3.00290 +- 0.00501 seconds time elapsed ( +- 0.17% )
[1] https://lore.kernel.org/dwarves/20241128012341.4081072-1-ihor.solodrai@xxxxx/
[2] https://lore.kernel.org/dwarves/20241128012341.4081072-3-ihor.solodrai@xxxxx/
[3] https://lore.kernel.org/dwarves/20241128012341.4081072-10-ihor.solodrai@xxxxx/
Alan Maguire (2):
btf_encoder: simplify function encoding
btf_encoder: separate elf function, saved function representations
Ihor Solodrai (8):
dwarf_loader: introduce pre_load_module hook to conf_load
btf_encoder: introduce elf_functions struct type
btf_encoder: collect elf_functions in btf_encoder__pre_load_module
btf_encoder: switch to shared elf_functions table
btf_encoder: introduce btf_encoding_context
btf_encoder: remove skip_encoding_inconsistent_proto
dwarf_loader: introduce cu->id
dwarf_loader: multithreading with a job/worker model
btf_encoder.c | 639 +++++++++++++++++++++---------------
btf_encoder.h | 8 +-
btf_loader.c | 2 +-
ctf_loader.c | 2 +-
dwarf_loader.c | 352 ++++++++++++++------
dwarves.c | 44 ---
dwarves.h | 21 +-
pahole.c | 237 +++----------
pdwtags.c | 3 +-
pfunct.c | 3 +-
tests/reproducible_build.sh | 5 +-
11 files changed, 685 insertions(+), 631 deletions(-)
--
2.47.1
