The combination of the libunwind refactoring and build system changes
exposed a build-test regression for the libunwind build using a cached
feature dump. Re-send kwork fixes [1] moving a definition to kwork.h
to avoid breaking no libbpf builds.

1. Build Fix for libunwind:
   "make feature-dump" runs without LIBUNWIND=1, skipping the setup of
   libunwind LDFLAGS. When Makefile.feature is included to generate the
   cached FEATURE-DUMP, the local test-libunwind.bin compiles without
   arch-specific link flags (-lunwind-x86_64) and fails, caching a false
   negative ("feature-libunwind=0"). Later builds (such as make_libunwind_O
   reusing the dump) compile the unwinder based on CONFIG_LIBUNWIND (remote)
   but disable HAVE_LIBUNWIND_SUPPORT (local), breaking compilation due to
   missing maps__e_machine declarations in maps.h. Fixed by unconditionally
   setting up libunwind feature build flags in Makefile.config.

2. kwork ASAN Double-Free & Leaks:
   - Solves double-free in the record command argument array due to
     parse_options mutation, by preserving original pointers in a calloc'ed
     to_free buffer.
   - Ensures kwork_usage string is freed on all exit paths.

3. kwork Unified Memory Lifecycle:
   - Establishes a unified memory ownership model for kwork_work by
     ensuring work names are dynamically duplicated via strdup() and
     consistently freed using work_exit() and a central perf_kwork__exit()
     teardown path.

[1] https://lore.kernel.org/linux-perf-users/20260520190538.142018-32-irogers@google.com/

Ian Rogers (3):
  perf build: Unconditionally set up libunwind feature build flags
  perf kwork: Fix address sanitizer issues
  perf kwork: Fix memory management of kwork_work

 tools/perf/Makefile.config  | 103 ++++++++++++------------
 tools/perf/builtin-kwork.c  | 151 ++++++++++++++++++++++++++++--------
 tools/perf/util/bpf_kwork.c |  14 +++-
 tools/perf/util/kwork.h     |   9 +++
 4 files changed, 188 insertions(+), 89 deletions(-)

-- 
2.54.0.746.g67dd491aae-goog

On Thu, May 21, 2026 at 12:24 AM Ian Rogers <irogers@google.com> wrote:
>
> The combination of the libunwind refactoring and build system changes
> exposed a build-test regression for the libunwind build using a cached
> feature dump. Re-send kwork fixes [1] moving a definition to kwork.h
> to avoid breaking no libbpf builds.
>
> 1. Build Fix for libunwind:
>    "make feature-dump" runs without LIBUNWIND=1, skipping the setup of
>    libunwind LDFLAGS. When Makefile.feature is included to generate the
>    cached FEATURE-DUMP, the local test-libunwind.bin compiles without
>    arch-specific link flags (-lunwind-x86_64) and fails, caching a false
>    negative ("feature-libunwind=0"). Later builds (such as make_libunwind_O
>    reusing the dump) compile the unwinder based on CONFIG_LIBUNWIND (remote)
>    but disable HAVE_LIBUNWIND_SUPPORT (local), breaking compilation due to
>    missing maps__e_machine declarations in maps.h. Fixed by unconditionally
>    setting up libunwind feature build flags in Makefile.config.
>
> 2. kwork ASAN Double-Free & Leaks:
>    - Solves double-free in the record command argument array due to
>      parse_options mutation, by preserving original pointers in a calloc'ed
>      to_free buffer.
>    - Ensures kwork_usage string is freed on all exit paths.
>
> 3. kwork Unified Memory Lifecycle:
>    - Establishes a unified memory ownership model for kwork_work by
>      ensuring work names are dynamically duplicated via strdup() and
>      consistently freed using work_exit() and a central perf_kwork__exit()
>      teardown path.

It would be nice to land these fixes. Thomas reported the kwork crashes in:
https://lore.kernel.org/linux-perf-users/314df838-4c38-4f03-9515-ae1dabd09a54@linux.ibm.com/

Thanks,
Ian

> [1] https://lore.kernel.org/linux-perf-users/20260520190538.142018-32-irogers@google.com/
>
> Ian Rogers (3):
>   perf build: Unconditionally set up libunwind feature build flags
>   perf kwork: Fix address sanitizer issues
>   perf kwork: Fix memory management of kwork_work
>
>  tools/perf/Makefile.config  | 103 ++++++++++++------------
>  tools/perf/builtin-kwork.c  | 151 ++++++++++++++++++++++++++++--------
>  tools/perf/util/bpf_kwork.c |  14 +++-
>  tools/perf/util/kwork.h     |   9 +++
>  4 files changed, 188 insertions(+), 89 deletions(-)
>
> --
> 2.54.0.746.g67dd491aae-goog
>

On Fri, May 22, 2026 at 02:33:05PM -0700, Ian Rogers wrote:
> On Thu, May 21, 2026 at 12:24 AM Ian Rogers <irogers@google.com> wrote:
> >
> > The combination of the libunwind refactoring and build system changes
> > exposed a build-test regression for the libunwind build using a cached
> > feature dump. Re-send kwork fixes [1] moving a definition to kwork.h
> > to avoid breaking no libbpf builds.
> >
> > 1. Build Fix for libunwind:
> >    "make feature-dump" runs without LIBUNWIND=1, skipping the setup of
> >    libunwind LDFLAGS. When Makefile.feature is included to generate the
> >    cached FEATURE-DUMP, the local test-libunwind.bin compiles without
> >    arch-specific link flags (-lunwind-x86_64) and fails, caching a false
> >    negative ("feature-libunwind=0"). Later builds (such as make_libunwind_O
> >    reusing the dump) compile the unwinder based on CONFIG_LIBUNWIND (remote)
> >    but disable HAVE_LIBUNWIND_SUPPORT (local), breaking compilation due to
> >    missing maps__e_machine declarations in maps.h. Fixed by unconditionally
> >    setting up libunwind feature build flags in Makefile.config.
> >
> > 2. kwork ASAN Double-Free & Leaks:
> >    - Solves double-free in the record command argument array due to
> >      parse_options mutation, by preserving original pointers in a calloc'ed
> >      to_free buffer.
> >    - Ensures kwork_usage string is freed on all exit paths.
> >
> > 3. kwork Unified Memory Lifecycle:
> >    - Establishes a unified memory ownership model for kwork_work by
> >      ensuring work names are dynamically duplicated via strdup() and
> >      consistently freed using work_exit() and a central perf_kwork__exit()
> >      teardown path.
> 
> It would be nice to land these fixes. Thomas reported the kwork crashes in:
> https://lore.kernel.org/linux-perf-users/314df838-4c38-4f03-9515-ae1dabd09a54@linux.ibm.com/

Thanks, applied to perf-tools-next, for v7.2.

- Arnaldo