Messages in this thread | | | Date | Wed, 1 Oct 2008 11:08:09 +0200 | From | Ingo Molnar <> | Subject | Re: [PATCH 05/31] cpumask: Provide new cpumask API |
| |
* Mike Travis <travis@sgi.com> wrote:
> Mike Travis wrote: > > Ingo Molnar wrote: > ... > > > >> In what way will Rusty's changes differ? Since you incorporate some of > >> Rusty's changes already, could you please iterate towards a single > >> patchset which we could then start testing? > > > > Our timezones are not very conducive to a lot of email exchanges > > (and he's moving.) From what I've seen I believe he's leaning > > towards using struct cpumask * and less trickery than I have.
actually, that's quite sane to do. const_cpumask_t looked a bit weird to me.
the extra indirection to a cpumask_t is not a big issue IMO, so in that sense whether we pass by value or pass by reference is not a _big_ performance item.
The complications (both present and expected ones) all come from the allocations.
> Oh yeah, I forgot the other major point of Rusty's approach. He wants > the patchset to be completely bisectable. That's far from true in my > version.
well, it should be a smooth transition and completely bisectable, there's hundreds of usages of cpumask_t and quite many in the pipeline. It's far easier to _you_ to get this stuff to work if it's all gradual and is expected to work all across. Have a default-off debug mode that turns off compatible cpumask_t perhaps - we can remove that later on.
with 'struct cpumask' we could keep cpumask_t as the compatible API, and could see the impact of these changes in a very finegrained and gradual way. Seems like a fundamentally sane approach to me ...
Ingo
| |