Re: [PATCH printk v2 3/4] printk: Skip unfinalized records in panic

[Petr Mladek]

On Wed 2023-10-18 17:56:52, John Ogness wrote:
> On 2023-10-18, Petr Mladek <pmladek@xxxxxxxx> wrote:
> > So it is the _last_ finalized id from the timing POV. If there are
> > more CPUs storing and finalizing the messages in parallel then
> > it might change forth and back. There might be earlier non-finalized
> > records and newer finalized ones.
> >
> > It means that prb_next_seq() really is the best effort and
> > the description is not valid:
> 
> Well, the description was valid until prb_next_seq() was optimized and
> converted to best-effort with:
> 
> commit f244b4dc53e5 ("printk: ringbuffer: Improve prb_next_seq() performance")
> 
> > It would be great to document these subtle details especially when
> > we are going to depend on them.
> 
> Going through the various call sites of prb_next_seq(), I would argue
> that the above optimization introduced some bugs. I will investigate if
> prb_next_seq() can be fixed to match its description because the current
> users already depend on that.

I thought about caching the last seq returned by
prb_next_seq() instead of caching the last finalized record.

Also I thought about the highest sequence number accessed
by _prb_read_valid(). But it can be done lockless way only
on 64-bit systems. Well, it might be good enough. I doubt that
there are big 32-bit systems.

> WRT to this series, I have put together an alternative implementation
> that does not use prb_next_seq().

Great.

Best Regards,
Petr