Em Fri, May 26, 2023 at 08:43:18PM -0700, Ian Rogers escreveu:
> Following on cleaning up .data and .bss in:
> https://lore.kernel.org/lkml/20230526183401.2326121-1-irogers@google.com/
> Look for some probably too large stack allocations with -Wstack-usage=20000
> and pahole.
Thanks, applied.
- Arnaldo
> Don't attempt to cleanup variable length arrays like in:
> ```
> util/header.c: In function ‘write_cache’:
> util/header.c:1269:12: warning: stack usage might be unbounded [-Wstack-usage=]
> 1269 | static int write_cache(struct feat_fd *ff,
> | ^~~~~~~~~~~
> ```
>
> Also leave two allocations relating to session/event processing:
> ```
> util/auxtrace.c: In function ‘auxtrace_queues__add_indexed_event’:
> util/auxtrace.c:424:12: warning: stack usage is 65616 bytes [-Wstack-usage=]
> 424 | static int auxtrace_queues__add_indexed_event(struct auxtrace_queues *queues,
> | ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
> util/session.c: In function ‘perf_session__peek_events’:
> util/session.c:1822:5: warning: stack usage is 65648 bytes [-Wstack-usage=]
> 1822 | int perf_session__peek_events(struct perf_session *session, u64 offset,
> | ^~~~~~~~~~~~~~~~~~~~~~~~~
> ```
>
> The biggest win is for perf inject where 128kb becomes lazily
> allocated when aux or guest data is encountered.
>
> Ian Rogers (4):
> perf sched: Avoid large stack allocations
> perf script: Remove some large stack allocations
> perf inject: Lazily allocate event_copy
> perf inject: Lazily allocate guest_event event_buf
>
> tools/perf/builtin-inject.c | 31 +++++++++++++++++++++++++------
> tools/perf/builtin-sched.c | 26 ++++++++++++++++++++++----
> tools/perf/builtin-script.c | 17 +++++++++++++----
> 3 files changed, 60 insertions(+), 14 deletions(-)
>
> --
> 2.41.0.rc0.172.g3f132b7071-goog
>
--
- Arnaldo