Re: [PATCH 3/3] ring_buffer: Use try_cmpxchg instead of cmpxchg

[Joel Fernandes]

On Wed, Mar 1, 2023 at 10:49 AM Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
>
> On Wed, Mar 1, 2023 at 4:37 AM Uros Bizjak <ubizjak@xxxxxxxxx> wrote:
> >
> > On Tue, Feb 28, 2023 at 10:43 PM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
> > >
> > > On Tue, 28 Feb 2023 18:59:29 +0100
> > > Uros Bizjak <ubizjak@xxxxxxxxx> wrote:
> > >
> > > > Use try_cmpxchg instead of cmpxchg (*ptr, old, new) == old.
> > > > x86 CMPXCHG instruction returns success in ZF flag, so this change
> > > > saves a compare after cmpxchg (and related move instruction in
> > > > front of cmpxchg).
> > > >
> > > > Also, try_cmpxchg implicitly assigns old *ptr value to "old" when cmpxchg
> > > > fails. There is no need to re-read the value in the loop.
> > > >
> > > > No functional change intended.
> > >
> > > As I mentioned in the RCU thread, I have issues with some of the changes
> > > here.
> > >
> > > >
> > > > Cc: Steven Rostedt <rostedt@xxxxxxxxxxx>
> > > > Cc: Masami Hiramatsu <mhiramat@xxxxxxxxxx>
> > > > Signed-off-by: Uros Bizjak <ubizjak@xxxxxxxxx>
> > > > ---
> > > >  kernel/trace/ring_buffer.c | 20 ++++++++------------
> > > >  1 file changed, 8 insertions(+), 12 deletions(-)
> > > >
> > > > diff --git a/kernel/trace/ring_buffer.c b/kernel/trace/ring_buffer.c
> > > > index 4188af7d4cfe..8f0ef7d12ddd 100644
> > > > --- a/kernel/trace/ring_buffer.c
> > > > +++ b/kernel/trace/ring_buffer.c
> > > > @@ -1493,14 +1493,11 @@ static bool rb_head_page_replace(struct buffer_page *old,
> > > >  {
> > > >       unsigned long *ptr = (unsigned long *)&old->list.prev->next;
> > > >       unsigned long val;
> > > > -     unsigned long ret;
> > > >
> > > >       val = *ptr & ~RB_FLAG_MASK;
> > > >       val |= RB_PAGE_HEAD;
> > > >
> > > > -     ret = cmpxchg(ptr, val, (unsigned long)&new->list);
> > > > -
> > > > -     return ret == val;
> > > > +     return try_cmpxchg(ptr, &val, (unsigned long)&new->list);
> > >
> > > No, val should not be updated.
> >
> > Please see the definition of try_cmpxchg. The definition is written in
> > such a way that benefits loops as well as linear code and in the later
> > case depends on the compiler to eliminate assignment to val as a dead
> > assignment.
> >
> > The above change was done under the assumption that val is unused
> > after try_cmpxchg, and can be considered as a temporary
> > [Alternatively, the value could be copied to a local temporary and a
> > pointer to this local temporary could be passed to try_cmpxchg
> > instead. Compiler is smart enough to eliminate the assignment in any
> > case.]

Ah I need to be more careful how I type.

> If I understood Steve correctly, I think the "misleading" part is that
> you are passing a variable by address to try_cmpxchg() but not really
> modifying it (unlike in the loop patterns).

It does modify it, but I meant it does not use it.

> Perhaps it is more meaningful to have an API that looks like:
> bool cmpxchg_succeeded(TYPE ptr, TYPE old, TYPE new)
> Where old is not a pointer (unlike try_cmpxchg), and the API returns bool.
>
> Both cleaner to read and has the optimization you want, I believe.
>
> However, the other point is, this is useful only for slow paths, but

Useful only for fast paths...

> at least cmpxchg_succeeded() is more readable and less "misleading"
> than try_cmpxchg() IMO.
>

Proofreading emails properly from here on! Not after the fact!

 - Joel