Messages in this thread | | | Date | Tue, 15 Apr 2008 19:15:11 +0200 | From | Andi Kleen <> | Subject | Re: [PATCH] Replace completions with semaphores |
| |
Linus Torvalds wrote: > > On Tue, 15 Apr 2008, Andi Kleen wrote: >>> - probably add support for completions to do counting >> But that's just a semaphore, isn't it? > > Exactly. But the point here is: > > - nobody should use semaphores anyway (use mutexes)
For normal locks. But if you have N number of outstanding events you need to wait for the semaphore is the right primitive.
And it seems there is a not high but non trivial number of places in the kernel who have a legitimate need for this.
> - making *more* code use semaphores is wrong > - completions have a different _mental_ model > > IOW, this is not about implementation issues. It's about how you think > about the operations.
Ok so you just want to rename it.
Fine for me. I always found up() and down() unintuitive anyways (but it's admittedly better than "P" and "V" which some other systems use)
> We should _not_ implement completions as semaphores, simply because we > want to get *rid* of semaphores some day. > > So rather than this long and involved patch series that first makes > semaphores generic, and then makes them be used as completions, I'd much > rather just skip this whole pointless exercise entirely. > > Why have "generic semaphores" at all, if we want to get rid of them?
Because we still "counted completions" for some things and that's the same code?
Rather i suspect the real problem is not the name, but just not sure it gets abused. That is largely more a review problem and as far as I can figure out basically all the usual reviewers take care of that anyways. But renaming it also probably wouldn't hurt.
[IMHO I always thought we should have a maintained single "list of things for reviewers to watch out for" list somewhere]
-Andi
| |