On Thu, 13 Nov 2025 17:07:39 +0100
Sebastian Andrzej Siewior <[email protected]> wrote:

> On 2025-11-13 10:51:06 [-0500], Steven Rostedt wrote:
> > Yes, because they are only tested in sched_switch and fork and exit 
> > tracepoints.
> > 
> > Although, this was written when tracepoint callbacks were always called
> > under preempt disable. Perhaps we need to change that call to:
> > 
> >     tracepoint_synchronize_unregister()
> > 
> > Or add a preempt_disable() around the callers.  
> 
> Please don't. Please do a regular rcu_read_lock() ;)
> I tried to get rid of the preempt_disable() around tracepoints so that
> the attached BPF callbacks are not invoked with disabled preemption. I
> haven't followed up here in a while but I think Paul's SRCU work goes
> in the right direction.

I meant just reading the pid lists, which are usually called from
tracepoints that are in preempt_disabled locations.

Anyway, I can add rcu_read_lock() around the callers of it.

> 
> > I'm very nervous about using RCU here. It will add a lot more corner cases
> > that needs to be accounted for. The complexity doesn't appear to be worth
> > it. I'd rather just keep the raw spin locks than to convert it to RCU.
> > 
> > The seqcount makes sense to me. It's simple and keeps the same paradigm as
> > what we have. What's wrong with using it?  
> 
> I'm fine with it once you explained under what conditions retry can
> happen.  Thank you.

Thanks,

-- Steve

Reply via email to