Re: [PATCH 0/4] perf: Fix leaked events when sigtrap = 1

[Frederic Weisbecker]

Le Fri, Mar 29, 2024 at 08:23:49PM -0700, Ian Rogers a écrit :
> On Fri, Mar 29, 2024 at 4:58 PM Frederic Weisbecker <frederic@xxxxxxxxxx> wrote:
> >
> > While looking at task_work users I just noticed that perf doesn't flush
> > its own upon event exiting. This looks especially problematic with child
> > events. Please have a thourough look at the last patch, I may easily
> > have missed something within the maze.
> >
> > Frederic Weisbecker (4):
> >   task_work: s/task_work_cancel()/task_work_cancel_func()/
> >   task_work: Introduce task_work_cancel() again
> >   perf: Fix event leak upon exit
> >   perf: Fix event leak upon exec and file release
> >
> >  include/linux/perf_event.h |  1 +
> >  include/linux/task_work.h  |  3 ++-
> >  kernel/events/core.c       | 40 +++++++++++++++++++++++++++++++-------
> >  kernel/irq/manage.c        |  2 +-
> >  kernel/task_work.c         | 34 +++++++++++++++++++++++++++-----
> >  security/keys/keyctl.c     |  2 +-
> >  6 files changed, 67 insertions(+), 15 deletions(-)
> 
> Thanks for this! I wonder if this relates to fuzzing failures like:
> https://lore.kernel.org/linux-perf-users/CAP-5=fUa+-Tj2b_hxk96Qg5=Qu7jYHgHREbsmBa2ZmuF-X9QaA@xxxxxxxxxxxxxx/
> "[ 2519.138665] unexpected event refcount: 2; ptr=000000009c56b097"

Probably not since those seem to happen on perf_event_open() failures. This
looks different.

Thanks.


> 
> Thanks,
> Ian
> 
> > --
> > 2.44.0
> >