On Tue, Jun 22, 2021 at 10:42 AM Arnaldo Carvalho de Melo
[off-list ref] wrote:

Em Tue, Jun 22, 2021 at 09:33:23AM -0700, Ian Rogers escreveu:

quoted

On Tue, Jun 22, 2021 at 12:44 AM Riccardo Mancini [off-list ref] wrote:

quoted

Hi,

thanks for your comments.

On Mon, 2021-06-21 at 22:14 -0700, Ian Rogers wrote:

quoted

On Mon, Jun 21, 2021 at 4:44 PM Riccardo Mancini [off-list ref] wrote:

quoted

ASan reports a memory leak related to session->evlist never being deleted.
The evlist member is not deleted in perf_session__delete, so it should be
deleted separately.
This patch adds the missing deletion in perf-script.

Signed-off-by: Riccardo Mancini <redacted>
---
 tools/perf/builtin-script.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tools/perf/builtin-script.c b/tools/perf/builtin-script.c
index 1280cbfad4db..635a1d9cfc88 100644
--- a/tools/perf/builtin-script.c
+++ b/tools/perf/builtin-script.c

@@ -3991,7 +3991,7 @@ int cmd_script(int argc, const char **argv)
                zfree(&script.ptime_range);
        }

-       evlist__free_stats(session->evlist);

Should this be removed?

Probably not. I originally thought this was already taken care of by
evlist__delete, but it's not.
Oddly, this issue is not causing a memory leak in my simple test.

quoted

quoted

+       evlist__delete(session->evlist);

This looks like a bug, if it is a 'session' member, its a session method
that should delete it, probably perf_session__delete().

quoted

quoted

quoted

If the perf session "owns" the evlist, would it be cleaner to add this
to perf_session__delete?

I thought about that too, but that's not always true.
E.g., in perf-record, __cmd_record calls perf_session__delete,then cmd_record
calls evlist__delete on rec->evlist, which points to the same location to which
session->evlist pointed.

Agreed. I find it hard to understand the ownership properties in the
perf code. The missing delete is an example of the owner of the evlist
(the caller) not "knowing" it needed cleaning up. I'd like it if we
documented things like perf_sessions' evlist to say not owned, user
must clean up. The makes it unambiguous who has to take
responsibility. Having things clean up after themselves is of course
easiest, hence wanting this to be in perf_session__delete.

This specific case, from just reading the description on this message,
looks just like a bug/thinko.

Ack. Definitely worth merging the change. I think this is about the
7th address sanitizer bug Riccardo has fixed. Namhyung, Numfor, Luke
and myself have also contributed similar fixes. We set up some fuzz
testing on libtraceevent and there are currently 12 issues we've found
there. The nice thing with sparse compared to address sanitizer is the
compiler will point at the problem, you don't need to trigger an issue
in a test. There are some complicated ownership rules in session and
also in the reference counting issues that Riccardo has raised, so
perhaps there's scope for some more comments or other tidy ups.

Thanks,
Ian

quoted

Fwiw, I've been reading around things like sparse [1, 2] and Clang's
similar analysis [3] that people have looked to use like sparse [4]. I
don't see anything that handles memory allocation lifetimes, but
perhaps something will feed into C's standards by way of C++ [5].
Perhaps people have ideas to rewrite in checked C or Rust :-)

Some thoughts:
1) we can't have C++ as we're trying to follow kernel conventions [6]
2) we can't annotate code for things like sparse or thread safety
analysis, as checking for memory errors is out of scope for them, the
annotations don't exist, etc.
3) we can add comments, document the rules around pointers, perhaps
even invent empty annotations that may one day help with automated
checking.
4) we can try to clean up the ownership model to make bugs less likely.

I've heard concerns on non-kernel projects about annotation litter and
comments adding to complexity. I think your patch is good, it follows
the existing conventions. I wonder if we can learn something from the
fact the code was wrong to make it less likely we have wrong code in
the future. I'd be interested to hear what others think.

Thanks,
Ian

[1] https://lore.kernel.org/lkml/Pine.LNX.4.58.0410302005270.28839@ppc970.osdl.org/ (local)
[2] https://lwn.net/Articles/689907/
[3] https://clang.llvm.org/docs/ThreadSafetyAnalysis.html
[4] https://www.openwall.com/lists/kernel-hardening/2019/05/20/3
[5] https://github.com/isocpp/CppCoreGuidelines/blob/master/docs/Lifetime.pdf
[6] even concatenating a string is error prone in C :-(
https://lore.kernel.org/lkml/YMzOpgZPJeC2jGKf@kernel.org/ (local)

quoted

Thanks,
Riccardo

quoted

Thanks,
Ian

quoted

        perf_session__delete(session);

        if (script_started)
--
2.31.1

--

- Arnaldo