Re: [PATCH] ring-buffer: Have the buffer update counter be atomic

[Steven Rostedt]

On Fri, 11 Oct 2024 10:05:51 +0200
Petr Pavlu <petr.pavlu@xxxxxxxx> wrote:

> On 10/11/24 01:58, Steven Rostedt wrote:
> > From: Steven Rostedt <rostedt@xxxxxxxxxxx>
> > 
> > In order to prevent any subtle races with the buffer update counter,
> > change it to an atomic_t. Also, since atomic_t is 32 bits, move its
> > location in the ring_buffer_per_cpu structure next to "current_context" as
> > that too is only 32 bits (making 64 bit alignment).
> > 
> > The counter is only used to detect that the buffer has been updated when
> > the buffer verifier check is being done. It's not really that important
> > that it's atomic or not. But since the updates to the counter are never in
> > the fast path, having it be consistent isn't a bad thing.
> > 
> > Signed-off-by: Steven Rostedt (Google) <rostedt@xxxxxxxxxxx>
> > ---
> > Note, this is based on top of:
> > 
> >   https://lore.kernel.org/linux-trace-kernel/20240715145141.5528-1-petr.pavlu@xxxxxxxx/  
> 
> Sorry for not replying to your last comment on my patch, I was ill.
> 
> The member ring_buffer_per_cpu.cnt is intended to be accessed under the
> reader_lock, same as the pages pointer which it is tied to, so this
> change shouldn't be strictly needed.
> 

Right, but there was one location that the cnt was updated outside the
lock. The one I commented on. But instead of adding a lock around it, I
decided to just make it an atomic. Then there would be no need for the
lock. Hmm, it still needs a memory barrier though. At least a
smp_mb__after_atomic().

-- Steve