LWN: Comments on "Data-type profiling for perf"

It's not in 6.8

Hi-Angel — Fri, 05 Apr 2024 12:01:32 +0000

Okay, I figured out what it's caused by: it's because I have `-flto` in default options and evidently there's some bug in `perf` that makes it break when that's defined. After removing flto I managed to compile it.

It's not in 6.8

Hi-Angel — Fri, 05 Apr 2024 11:50:57 +0000

FTR, apparently there's some problem with `perf` in 6.8. I was tired of waiting for 6.8 perf to appear in Arch repos so decided to download PKGBUILD and compile it myself. Well, now I know why it takes for long for perf to get updated: it's unbuildable, the linking stage fails with (also, Idk why text below is not getting aligned, I inserted indentation before it 🤷‍♂️):

LINK perf
/usr/bin/ld: /usr/lib/gcc/x86_64-pc-linux-gnu/13.2.1/../../../../lib/Scrt1.o: in function `_start':
(.text+0x1b): undefined reference to `main'
/usr/bin/ld: pmu-events/pmu-events-in.o: in function `map_for_pmu':
pmu-events.c:(.text+0x174): undefined reference to `perf_pmu__getcpuid'
/usr/bin/ld: pmu-events.c:(.text+0x1ac): undefined reference to `strcmp_cpuid_str'
/usr/bin/ld: pmu-events/pmu-events-in.o: in function `pmu_events_table__for_each_event':
(.text+0x311): undefined reference to `pmu__name_match'
/usr/bin/ld: pmu-events/pmu-events-in.o: in function `pmu_events_table__find_event':
(.text+0x4ce): undefined reference to `pmu__name_match'
/usr/bin/ld: (.text+0x5e0): undefined reference to `pmu__name_match'
/usr/bin/ld: pmu-events/pmu-events-in.o: in function `pmu_events_table__num_events':
(.text+0x6f7): undefined reference to `pmu__name_match'
/usr/bin/ld: pmu-events/pmu-events-in.o: in function `perf_pmu__find_events_table':
(.text+0xa33): undefined reference to `pmu__name_match'
/usr/bin/ld: pmu-events/pmu-events-in.o:(.text+0xae3): more undefined references to `pmu__name_match' follow
/usr/bin/ld: pmu-events/pmu-events-in.o: in function `find_core_events_table':
(.text+0xb79): undefined reference to `strcmp_cpuid_str'
/usr/bin/ld: pmu-events/pmu-events-in.o: in function `find_core_metrics_table':
(.text+0xc09): undefined reference to `strcmp_cpuid_str'
/usr/bin/ld: /home/constantine/Projects/builds/linux-tools/src/linux/tools/perf/libsymbol/libsymbol.a(libsymbol-in.o): in function `__tolower':
/home/constantine/Projects/builds/linux-tools/src/linux/tools/lib/symbol/kallsyms.c:52:(.text+0x10): undefined reference to `_ctype'
/usr/bin/ld: /home/constantine/Projects/builds/linux-tools/src/linux/tools/perf/libsymbol/libsymbol.a(libsymbol-in.o): in function `__toupper':
/home/constantine/Projects/builds/linux-tools/src/linux/tools/lib/symbol/kallsyms.c:59:(.text+0x3e): undefined reference to `_ctype'
collect2: error: ld returned 1 exit status

It's not in 6.8

Hi-Angel — Sat, 23 Mar 2024 09:40:23 +0000

Oh, actually, sorry for the confusion, it seems that for some reason Archlinux packages updated the kernel but still didn't update `perf` utility that goes with it, Idk why. It was marked outdated 12 days ago. Anyway, please disregard my comment, will re-test it once Arch update perf to the kernel version.

It's not in 6.8

Hi-Angel — Sat, 23 Mar 2024 09:36:50 +0000

> on its way to Linux v6.8 in January

I'm on 6.8.1 and calling a `perf annotate --data-type` results in a `Error: unknown option `data-type'`. So it didn't make it.

Data-type profiling for perf

rwmj — Thu, 04 Jan 2024 10:47:11 +0000

This looks fantastic. Next step would be some kind of latency analysis. I wonder if it's possible to see which fields have high latency for writes (which might indicate a cache line "ping-ponging" between cores)?

Data-type profiling for perf

taladar — Wed, 27 Dec 2023 10:14:09 +0000

If it is cheaper for the compiler to compute writing the information to a separate file as part of the compile process might also be an option.

Data-type profiling for perf

Wol — Wed, 27 Dec 2023 10:07:37 +0000

Sounds to me like a straightforward database file with index ...

Cheers,
Wol

Data-type profiling for perf

rywang014 — Tue, 26 Dec 2023 21:27:23 +0000

Can we do some large scale automations to discover more layout optimizations? It can run a wide range of benchmarks with this tool, and find if there are structs with multiple active cache lines and can be shuffled to a same cache line.

Data-type profiling for perf

DanilaBerezin — Tue, 26 Dec 2023 21:08:46 +0000

I think slower builds and bloated binaries are an okay trade off for a debug build. But in general, yeah I would agree, I think if it's possible to create a secondary program that inverts the lists after the build, that would probably be preferable.

Data-type profiling for perf

dankamongmen — Sun, 24 Dec 2023 00:16:02 +0000

this looks absolutely outstanding

Data-type profiling for perf

roc — Sat, 23 Dec 2023 05:15:19 +0000

We don't really want compilers to emit redundant DWARF tables. That slows down builds and creates bloated binaries. A better approach would be to have a tool that can build inverted location lists from the regular location lists, persistently caching the results by build-ID when that's helpful.

Data-type profiling for perf

acme — Fri, 22 Dec 2023 12:12:48 +0000

It's in https://git.kernel.org/pub/scm/linux/kernel/git/perf/perf..., probably will move to perf-tools-next later today, on its way to Linux v6.8 in January.

Great article! Should encourage people to test it and help with finding issues, fixing problems and adding more features.

Data-type profiling for perf

Sesse — Fri, 22 Dec 2023 08:11:05 +0000

Oh wow, I've been wanting this for some time. At some point, I even wrote a hack for something similar myself :-)

Data-type profiling for perf

pctammela — Thu, 21 Dec 2023 16:51:34 +0000

This is really nice, great work!