Messages in this thread | | | Date | Wed, 1 Mar 2023 13:18:31 -0500 | From | Steven Rostedt <> | Subject | Re: [PATCH 3/3] ring_buffer: Use try_cmpxchg instead of cmpxchg |
| |
On Wed, 1 Mar 2023 18:16:04 +0100 Uros Bizjak <ubizjak@gmail.com> wrote:
> > It's different, but I'm not so sure it's beneficial. > > This is the place we are looking for. Please see that a move at $1464 > and a compare at $1473 is missing in the assembly from the patched > code. If it is beneficial ... well, we achieved the same result with > two instructions less in a loopy code.
Note, none of these locations are in fast paths. (now if we had a local_try_cmpxchg() then we could improve those locations!).
Thus my main concern here is for correctness. Unfortunately, to add a cmpxchg_success() would require a lot of work, as all the cmpxchg() functions are really macros that do a lot of magic (the include/linux/atomic/atomic-instrumented.h says its even generated!). So that will likely not happen.
I have mixed feelings for this patch. I like the fact that you are looking in optimizing the code, but I'm also concerned that it could cause issues later down the road.
I am leaning to just taking this, and hope that it doesn't cause problems. And I would really love to change all the local_cmpxchg() to local_try_cmpxchg(). Hmm, I wonder how much of an effort that would be to implement those?
-- Steve
| |