Re: [PATCH v4 0/6] perf: Make SIGTRAP and __perf_pending_irq() work on RT.
From: Marco Elver
Date: Tue Jun 25 2024 - 09:43:42 EST
On Mon, 24 Jun 2024 at 17:27, Sebastian Andrzej Siewior
<bigeasy@xxxxxxxxxxxxx> wrote:
>
> Hi,
>
> Arnaldo reported that "perf test sigtrap" fails on PREEMPT_RT. Sending
> the signal gets delayed until event_sched_out() which then uses
> task_work_add() for its delivery. This breaks on PREEMPT_RT because the
> signal is delivered with disabled preemption.
>
> While looking at this, I also stumbled upon __perf_pending_irq() which
> requires disabled interrupts but this is not the case on PREEMPT_RT.
>
> This series aim to address both issues while not introducing a new issue
> at the same time ;)
> Any testing is appreciated.
Tested-by: Marco Elver <elver@xxxxxxxxxx>
Ran lots of concurrent copies of the "sigtrap_threads" and
"remove_on_exec" tests (with lockdep on), and it all survived. Fuzzer
is still running but hasn't found anything relevant yet.
> v3…v4: https://lore.kernel.org/all/20240322065208.60456-1-bigeasy@xxxxxxxxxxxxx/
> - Rebased on top of Frederic's series
> (https://lore.kernel.org/all/20240621091601.18227-1-frederic@xxxxxxxxxx)
> - Frederick pointed out that perf_pending_task() needs to
> perf_swevent_get_recursion_context() in order not to recurse if
> something within perf_swevent_.*_recursion_context() triggers a
> software event. To address this, the counters have been moved to
> the task_struct (#3 + #4) and preemt_disable() has been replaced
> with a RCU-read lock (#5).
> - The remaning logic same that means the event is pushed to task-work
> instead of delivering from IRQ-work. The series was tested with
> remove_on_exec as suggested by Marco Elver: On PREEMPT_RT a single
> invocation passes, 100 parallel invocations report (for some)
> unexpected SIGTRAPs and timeouts. This also observed on !RT
> (without the series) with a higher task-count.
>
> v2…v3: https://lore.kernel.org/all/20240312180814.3373778-1-bigeasy@xxxxxxxxxxxxx/
> - Marco suggested to add a few comments
> - Added a comment to __perf_event_overflow() to explain why irq_work
> is raised in the in_nmi() case.
> - Added a comment to perf_event_exit_event() to explain why the
> pending event is deleted.
>
> v1…v2: https://lore.kernel.org/all/20240308175810.2894694-1-bigeasy@xxxxxxxxxxxxx/
> - Marco pointed me to the testsuite that showed two problems:
> - Delayed task_work from NMI / missing events.
> Fixed by triggering dummy irq_work to enforce an interrupt for
> the exit-to-userland path which checks task_work
> - Increased ref-count on clean up/ during exec.
> Mostly addressed by the former change. There is still a window
> if the NMI occurs during execve(). This is addressed by removing
> the task_work before free_event().
> The testsuite (remove_on_exec) fails sometimes if the event/
> SIGTRAP is sent before the sighandler is installed.
>
> Sebastian
>