Re: [GIT PULL] sigqueue cache fix

From: Thomas Gleixner
Date: Wed Jul 07 2021 - 05:47:54 EST


On Mon, Jun 28 2021 at 12:02, Linus Torvalds wrote:
> And as the whole - and only - point of the sigqueue cache was a very
> subtle performance and latency issue, I don't think we want to use
> locks or atomics. It's why my revert commit suggests re-purposing the
> "percpu_cmpxchg()" functionality: that would likely be a good model at
> least for x86 and arm.
> But while we have "percpu_cmpxchg()", I don't think we currently don't
> really have that kind of operation where it
> (a) works on a non-percpu variable (but we can force casts, I guess)
> (b) has "acquire" semantics
> We do have the *atomic* cmpxchg with acquire semantics, but atomics
> are rather more expensive than we'd probably really want.

first of all, sorry for being non-responsive on this. I'll have a fresh
look at this with your comments in mind and be more careful.