Re: [PATCH 1/2] tracing: Simplify and fix "buffered event" synchronization

[Steven Rostedt]

On Fri, 1 Dec 2023 15:17:35 +0100
Petr Pavlu <petr.pavlu@xxxxxxxx> wrote:

> Ok, keeping the current approach, my plan for v2 is to prepare the
> following patches:
> 
> * Fix for the missing increment+decrement of trace_buffered_event_cnt
>   on the current CPU in trace_buffered_event_disable().
> 
>   Replace smp_call_function_many() with on_each_cpu_mask() in
>   trace_buffered_event_disable(). The on_each_cpu_mask() function has
>   also an advantage that it itself disables preemption so doing that can
>   be then removed from trace_buffered_event_disable().

OK.

> 
> * Fix the potential race between trace_buffered_event_enable() and
>   trace_event_buffer_lock_reserve() where the latter might already see
>   a valid trace_buffered_event pointer but not all initialization yet.
> 
>   I think this might be actually best to address by using the same
>   maintenance exclusion as is implemented in
>   trace_buffered_event_disable(). It would make both maintenance
>   operations consistent but for the cost of making the enable operation
>   somewhat slower.

I wouldn't do them the same just to make them consistent. I think the
smp_wmb() is sufficient. Don't you think?

> 
> * Fix the WARN_ON_ONCE(!trace_buffered_event_ref) issued in
>   trace_buffered_event_disable() when trace_buffered_event_enable()
>   previously fails.
> 
>   Add a variable/flag tracking whether trace_buffered_event is currently
>   allocated and use that for driving if a new allocation needs to be
>   done when trace_buffered_event_enable() is called, or the buffers
>   should be really freed when trace_buffered_event_disable() is invoked.
> 
>   Not sure if the mentioned alternative of leaving trace_buffered_event
>   partially initialized on failure is preferred instead.

I do not really have a preference for either solution. They both are bad if
it happens ;-)

> 
> * Fix the potential race between trace_buffered_event_disable() and
>   trace_event_buffer_lock_reserve() where the latter might still grab
>   a pointer from trace_buffered_event that is being freed.
> 
>   Replace smp_wmb() with synchronize_rcu() in
>   trace_buffered_event_disable().

Sounds good.

Thanks!

-- Steve